Category: Law Enforcement & Criminal Justice

Policing, sentencing, surveillance

  • The Vanishing Draft: How a Product Built to Forget, a Six-Hundred-Thousand-Report Milestone, and Two State Statutes Collided Over Police Accountability

    The Vanishing Draft: How a Product Built to Forget, a Six-Hundred-Thousand-Report Milestone, and Two State Statutes Collided Over Police Accountability

    AI now drafts police reports in roughly 600 departments nationwide — while the leading tool was engineered to leave no record of what it wrote. Two enacted state laws, a wall of prosecutor refusals, and a November wave of study bills are forcing the question the design deferred: who wrote this report?


    The Scale of the Shift

    Axon Enterprise, the company best known for the Taser, says its AI report-writing tools are now in use by 600 police departments across the country and have helped write 600,000 police reports, saving an estimated 300,000 hours according to its customer surveys [1]. Its AI Era product bundle — which includes Draft One, the tool that writes incident narratives from body-camera audio — costs $199 per officer per month, and the company booked approximately $750 million in AI Era contracts in the year after Draft One launched, with first-quarter 2026 revenue from AI products growing more than 700 percent year-over-year [1]. These are no longer pilot programs; they are infrastructure.

    The stakes follow from what a police report is. Officers spend as much as 15 hours a week on reports, and those documents are either stored for reference or handed to prosecutors who decide whether to file charges — making them, as one criminologist put it, a crucial part of the criminal justice system rather than clerical residue [1]. When the drafting of that document moves from an officer’s memory to a language model’s output, the policy question is not whether paperwork gets faster. It is whether the justice system retains the ability to ask, after the fact, what the machine said and what the human changed.


    Two Enacted Laws, Two Philosophies

    Only two states have enacted statutes governing AI-assisted police reports, and they anchor opposite ends of a spectrum: disclosure with full audit infrastructure, versus disclosure with agency self-governance.

    • California SB 524 (Penal Code §13663). Signed October 10, 2025 and effective January 1, 2026, the law requires every law enforcement agency to maintain a policy under which an official report generated fully or partially with AI must, on each page or in the body of the text, identify every specific AI program used and prominently state: “This report was written either fully or in part using artificial intelligence” [2]. The officer must sign the report, verifying they reviewed its contents and that “the facts contained in the official report are true and correct” [2]. The core logic is evidentiary: only the final signed report counts as the officer’s statement, and the first AI-generated draft must be retained for as long as the final report — so that a court, a defense attorney, or a records requester can always compare what the machine proposed against what the officer swore to.
    • Utah SB 180 (Law Enforcement Usage of Artificial Intelligence). Effective May 2025, the statute directs each agency to adopt a written policy governing AI use, requires officers to disclose when AI was involved in generating a report, and requires the officer’s certification that the report was reviewed for accuracy [3][4]. Utah’s approach puts the compliance burden on the agency’s own policy process rather than a statutory retention mandate — the state sets the floor of disclosure, and the department writes the procedure.

    California’s statute reaches one step further than Utah’s, and the step is vendor-facing. A contracted vendor “shall not share, sell, or otherwise use information provided by a law enforcement agency to be processed by artificial intelligence” except for the agency’s own purposes or pursuant to a court order, though vendors may access processed data for troubleshooting, bias mitigation, accuracy improvement, or system refinement [2]. An audit trail must identify, at minimum, the person who used the AI and the video and audio footage used to create the report [2]. The law effectively nationalizes, for California agencies, a data-governance standard that no vendor contract had established on its own.


    A Product Built to Forget

    The reason California’s retention mandate matters is that the dominant product was designed without it. An Electronic Frontier Foundation investigation — built on police reports, emails, procurement documents, department policies, and Axon’s own manuals and marketing materials — concluded that Draft One is engineered so that no one, including a police chief or an independent researcher, can meaningfully audit its use [5]. Draft One processes only the audio from body cameras, not the video; the officer edits the generated text and then copies it into the department’s records system, at which point the draft disappears when the browser tab closes [5]. There is no log distinguishing which portions of a final report were written by the machine and which by the officer, except the officer’s own recollection [5].

    EFF, which obtained a recording of an Axon roundtable discussion, quotes the company’s senior principal product manager for generative AI explaining the retention decision in terms that will likely be quoted in courtrooms for years:

    “So we don’t store the original draft and that’s by design and that’s really because the last thing we want to do is create more disclosure headaches for our customers and our attorney’s offices.” [5]

    The consequences surface in public-records responses. The Lafayette, Indiana Police Department — one of Axon’s most promoted Draft One clients — told EFF: “Regarding the attached request, we do not have the ability to create a list of reports created through Draft One. They are not searchable. This request is now closed.” [5] The Palm Beach County Sheriff’s Office, by contrast, requires a disclosure footer in every AI-assisted report and was able to isolate more than 3,000 Draft One reports generated between December 2024 and March 2025 with a simple keyword search [5]. The difference between the two agencies is not the software — it is whether local policy forced a marker into the text. Axon itself tracks which reports used Draft One internally, and provided a Colorado agency custom code to extract the data after a media request revealed the agency could not [5].

    Axon’s stated position is that officers remain in control: the company describes its approach as “officer-in-the-loop,” saying narratives “cannot be submitted without officer review and approval,” and a spokesperson noted that “AI capabilities and Draft One have evolved significantly in the last two years” [1]. EFF’s finding is narrower but harder to answer: even if officers review every draft, the system keeps no record proving they did, and Axon disclosed a bug that let officers bypass the review guardrails on at least three occasions [5]. A safeguard that leaves no trace is a safeguard that cannot be audited — which is precisely the gap SB 524’s first-draft retention requirement was written to close.


    The Accuracy Record

    The retention question would be academic if the drafts were reliably right. Public records obtained by Forbes suggest they are not, and that the editing burden can consume the promised savings. In Lafayette, officers testing Axon’s Form One tool — which fills in names, license plates, and vehicle details from camera audio — reported in emails that the tool struggled to record the right names or car plates even when they were clearly stated in the footage, with one officer writing: “I know it doesn’t save me time and I know it has inaccuracies that I will have to edit,” and another concluding that “Form One dramatically increases the time it takes to finish reports” — a form that took 30 seconds by hand now taking three minutes [1].

    The only data-driven study of the flagship tool points the same direction. Criminologist Ian Adams, a former police officer, studied the Manchester, New Hampshire Police Department’s use of Draft One for a 2024 Journal of Experimental Criminology paper and found no improvement in report completion time: officers without access filed at the same pace because Draft One users spent their time removing irrelevant information, fixing inaccuracies, and adding facts the system missed [1]. Manchester dropped the tool in 2024, the same year Anchorage police stopped using it citing zero time savings — yet about half of Manchester’s officers still believed it had helped, a gap Adams called “a definite gap between the perception and objective reality” [1]. That perception gap is itself a policy problem: if officers systematically overestimate the tool’s reliability, the human review that the whole framework depends on weakens precisely where it is hardest to observe.


    The Prosecutor Wall

    The fastest-moving constraint has come not from legislatures but from the officials who decide what evidence they will take to court. The Prosecuting Attorney’s Office in King County, Washington — which includes Seattle — informed law enforcement agencies in 2024 that it would not accept reports written with AI assistance because of hallucinations, and its policy remains in force [1]. Its 2024 memo was explicit:

    “We do not fear advances in technology — but we do have legitimate concerns about some of the products on the market now… For now, our office has made the decision not to accept any police narratives that were produced with the assistance of AI.” [1]

    King County Prosecutor Daniel Clark described the concrete failure modes: AI getting the names of witnesses and officers wrong, or placing officers at a scene when they were only on the radio. “It’s a type of error that can really challenge an officer’s credibility when they’re on the stand, and then that puts the officer at risk of being either called lazy, sloppy or dishonest,” Clark said. “None of those are good options.” [1] The logic cuts across the usual ideological lines: a prosecutor refusing AI-drafted reports is not protecting defendants from police so much as protecting the evidentiary value of police work itself.

    The refusals are stacking up beyond Seattle. Connecticut’s Chief State’s Attorney imposed a statewide, non-statutory moratorium on AI-drafted official reports in April 2026, and county-level prosecutor restrictions are in place in King County, Wayne County, Michigan, and Hays County, Texas [4]. No state has banned the practice by statute — a Nevada ban was stripped from its bill before it died in 2025, and Minnesota’s government-wide prohibition on generative AI in official records lapsed at the May 2026 adjournment [4]. The pattern is a governance inversion: the branch of the system with the least formal rulemaking power is setting the de facto national standard, one charging decision at a time.


    The Vendor Incentive Problem

    Disclosure laws address who wrote the report, but the ACLU has raised a second-order concern: the company that writes the draft may have reasons of its own to shape it. Axon and its rival Motorola Solutions do not merely sell report-writing AI — both sell license plate readers, surveillance cameras, body cameras, drones (in Axon’s case), Tasers, and real-time crime center software, meaning the same vendor’s products may be what the report is about [6]. A report-drafting model, in other words, could be trained or prompted to mention some technologies favorably, or not at all.

    The concern is not hypothetical. EFF found that Vigilant Solutions — the license-plate-reader company later acquired by Motorola — recommended in a training slideshow that officers omit its technology from reports when possible: “If the use of the LPR equipment can be left out of the report, that is what is recommended” [6]. The precedent shows a police-technology vendor actively coaching officers on report content that affected the vendor. California’s vendor-data restrictions and audit trail are the first statutory answer to that conflict of interest; they still do not reach the model itself — what it was fine-tuned to include, omit, or emphasize — which remains invisible to agencies, courts, and the public [6].


    The Next Wave: Study Bills and Deadlines

    The statehouses are now processing what the first two laws started. New Jersey is the most recent test case: two bills — A-759 on facial recognition and A-3076 on AI — would direct the attorney general to study how police use the technologies and recommend statewide policies, with A-3076 calling for a policy requiring officers to review AI-generated reports for accuracy and completeness [7]. Both cleared the Assembly Public Safety and Preparedness Committee on October 19 [7]. Assemblyman Balvir Singh, whose district already has departments using AI report writers without statewide guidance, named the core risk:

    “The concern is that police reports are often central to court proceedings, where specific words, phrasing, and details matter,” Singh said. “Generative AI doesn’t just clean up grammar. It can add supporting details, substitute words, and change phrasing for clarity and flow. These unintended changes, especially when they go unnoticed, can alter the meaning of an officer’s account and potentially affect legal proceedings.” [7]

    State Senator Benjie Wimberly, who is expected to introduce identical bills in the upper chamber, framed the choice in fewer words: “We shouldn’t have to wait for something to go wrong before establishing appropriate safeguards. Public safety and civil liberties are not mutually exclusive. We have an obligation to protect both.” [7] Skeptics argue studies are the slower of two paths — Montclair State justice studies professor Jason Williams would prefer accountability measures, use restrictions, and independent audits now, warning that “Officer review is necessary… but it may not be sufficient because sometimes people have this tendency to just rely on the AI output” [7].

    Elsewhere the calendar is doing the work:

    • New York. S10425/A9253-A would have required 90 days’ public notice before deploying a covered AI product and disclosure to the defense within 20 days of any post-arrest AI use; it stalled in committee and dies with the 2025–26 term in January 2027 unless reintroduced [4]. Its defense-disclosure clock is the most defendant-facing provision any legislature has drafted.
    • Virginia. HB 1294, requiring disclosure, first-draft retention, and an audit trail, was carried over to the 2027 session [4] — the closest pending analogue to California’s regime.
    • Connecticut, Michigan, Washington. Prosecutor-level restrictions remain the operative rule where legislatures have not acted, a reminder that the current national baseline is a patchwork of charging policies rather than law [4].

    What to Watch

    • The first enforcement test of SB 524. California’s first-draft retention requirement took effect January 1, 2026, and EFF’s finding is that Draft One “cannot comply” with it as designed [5]. Whether Axon ships a retention feature for California agencies — and whether any department without it continues using the tool — is the first real compliance question.
    • Disclosure in the courtroom. California requires the disclosure statement to be “readily apparent to the reader” on every page [2]. The first appellate fight over whether an AI-assisted report was properly disclosed to a jury will define how far the statute reaches in practice.
    • Whether prosecutors codify their refusals. The King County and Connecticut policies are internal decisions, reversible at will [4]. If a legislature converts any of them into statute, the de facto standard becomes binding.
    • The vendor audit trail. Axon tracks Draft One report usage internally that its customer agencies cannot extract [5]. Whether California’s audit-trail requirement forces that data into agency hands is an open implementation question — and a test of whether disclosure law can outrun product design.
    • New Jersey’s study reports. If A-3076-style provisions advance, the attorney general’s recommended statewide policy — particularly its officer-review requirement — would give the study-bill approach its first output [7].

    The collision at the center of this story is simple to state. A product that writes the foundational document of criminal prosecution was designed, in its vendor’s own words, to avoid “disclosure headaches” — and two state legislatures have now answered that design with a disclosure mandate. The question the next year will settle is whether the record of what the machine wrote becomes something the justice system can see, or whether six hundred thousand reports join the system with no one able to say which words were the officer’s.


    References

    [1] Forbes — [SECONDARY] Brewster, Thomas. “Axon Says AI Police Reports Save Time. Public Records Show They Get Facts Wrong.” July 22, 2026.

    [2] California Legislative Information — [PRIMARY] SB 524 (Arreguín), Chapter 587, Penal Code §13663. Approved October 10, 2025; effective January 1, 2026.

    [3] Utah State Legislature — [PRIMARY] SB 180, “Law Enforcement Usage of Artificial Intelligence.” 2025 General Session; effective May 2025.

    [4] Barricade AI — [SECONDARY] “AI Police Report Laws by State: 2026 Tracker.” Updated November 2026. (vendor-maintained research tracker)

    [5] Electronic Frontier Foundation — [ADVOCACY] Guariglia, Matthew, and Dave Maass. “Axon’s Draft One Is Designed to Defy Transparency.” July 10, 2025. (digital rights organization)

    [6] American Civil Liberties Union — [ADVOCACY] Stanley, Jay. “Another Problem With AI-Assisted Police Reports: Companies Can Have Incentives to Distort Their Content.” August 28, 2026. (civil liberties organization)

    [7] The Jersey Vindicator — [SECONDARY] Janoski, Steve. “NJ lawmakers want safeguards for police use of AI, facial recognition software.” November 2, 2026.


    AI automated and generated content. Human reviewed.


    Listen to the Podcast: AI Policy for the Bears

    Episode 45: The Vanishing Draft: How a Product Built to Forget, a Six-Hundred-Thousand-Report Milestone, and Two State Statutes Collided Over Police Accountability — click to listen

    Two bears break down this story on AI Policy for the Bears, Episode 45.

  • The Listener in the Wire: How Ohio Put a Million Dollars Behind AI Call Monitoring, a County Logged 88 Confessions in a Quarter, and the FCC Voted to Bill Inmates for Their Own Surveillance

    The Listener in the Wire: How Ohio Put a Million Dollars Behind AI Call Monitoring, a County Logged 88 Confessions in a Quarter, and the FCC Voted to Bill Inmates for Their Own Surveillance

    A state budget line, a sheriff’s quarterly numbers, and a federal rate change are quietly building the largest AI surveillance system aimed at incarcerated Americans — with the people being watched paying part of the bill.


    The Policy Moment

    Three developments, barely noticed outside criminal justice circles, converged in 2025 and 2026 to make AI monitoring of prison phone calls one of the fastest-moving AI deployments in American law enforcement. First, the FCC reversed a 2024 reform and voted to let telecom companies pass the cost of recording, transcribing, and AI-analyzing calls back onto incarcerated people and their families, effective April 6, 2026. Second, Securus Technologies — the country’s largest prison telecom — disclosed it trained a large language model on seven years of Texas prison calls and is now piloting it to predict crimes before they happen. Third, states are appropriating public money to scale the technology: Ohio’s recently passed state budget included $1 million for the Department of Rehabilitation and Correction to buy software that can “transcribe and analyze all inmate phone calls.” None of these moves required new legislation explicitly authorizing AI surveillance of prisons. That is precisely why they matter: the deployment is running ahead of any dedicated policy framework.


    What Ohio Did

    Ohio’s state budget provided $1 million for the Ohio Department of Rehabilitation and Correction to purchase software able to “transcribe and analyze all inmate phone calls,” with the resulting data made accessible to law enforcement agencies across the state to support criminal investigations. The program is funded as a pilot through 2026, and the rollout begins at Ross Correctional Institution in Chillicothe — the prison where, prosecutors say, an incarcerated man murdered a corrections officer on Christmas Day 2024, amid guard-union reports of understaffing and a near-tripling of use-of-force incidents between 2013 and 2023. [1]

    Two details in the Ohio story illustrate how AI policy actually gets made in this space — not through a standalone statute, but through budget riders and lobbying. The vendor whose software won the funding, Texas-based LeoTech, hired three lobbyists from two law firms to press its case to House lawmakers, and the ODRC director personally advocated for the appropriation. The supporting lawmakers’ logic is straightforward: state policy already allowed employees and contractors to monitor and record any call at random or “for cause,” but manual review was capped by staffing. AI simply removes the staffing constraint. As House Finance Chairman Brian Stewart put it, the appeal was “the potential to do what we’re already doing, in a more efficient way with less man hours and taxpayer expense.” [1]

    That logic is worth pausing on, because it is the core argument driving AI adoption across law enforcement. If the underlying power to surveil already exists, the argument runs, then automating it is merely an efficiency upgrade. But efficiency in surveillance is not neutral. The same dollar of monitoring budget that previously covered a sample of calls now covers all of them — which changes the character of the power, not just its speed. Legal scholars and advocates argue that this quantitative leap produces a qualitatively different system, what one Columbia Law lecturer described as a “permeating surveillance system” that sweeps in law-abiding people on the outside of the prison walls too. [1]


    What Cuyahoga County Says the Technology Does

    The clearest public window into how these systems perform in practice comes from Cuyahoga County, which has run the LeoTech Verus system on its Cleveland jail calls for the past year. County officials presented the case for a $2.7 million, two-year contract extension to council in September 2026, and their numbers are the first sustained body of evidence on an AI call-monitoring deployment in a major American jail. [2]

    • 88 confessions captured in one quarter. In the second quarter of 2026 alone, the AI-assisted monitoring system captured 88 confessions, according to the sheriff department’s chief deputy. The system flags calls by keyword, alerting staff in real time to potential criminal activity or self-harm. The confessions figure is the headline metric officials cite because it is the most directly verifiable outcome: an admission in a recorded call is evidence that can be checked against a case file, which is why it anchors the county’s pitch for continuing the contract. [2]
    • 23 homicides solved, 30 suicides prevented. Chief Deputy Aaron Reese told council the system has helped solve 23 homicides and prevent 30 suicides, and that the jail has recorded no suicide or overdose death since adoption. These numbers are self-reported by the department, not independently audited, but they represent the strongest empirical claim yet that AI monitoring saves lives behind bars. [2]
    • The privilege firewall is procedural, not technical. Attorney-client calls are protected not by the AI but by registration: lawyers sign up with jail officials and receive instructions on how to “safeguard” the calls, and the county says such calls are not recorded. Given that the same underlying provider, Securus, has paid settlements in other states for recording privileged calls, the county’s safeguard depends entirely on process discipline holding up under scale. [2]

    The Cuyahoga numbers show why this technology is politically durable: it produces countable successes with sympathetic victims (homicide victims, would-be suicides) and low visibility for its costs. The 88 confessions are named and counted; the calls never made because a family could no longer afford the per-minute rate are not. That asymmetry — measurable benefits, invisible costs — is the structural bias any regulatory response has to overcome.


    The Securus Model: Training AI on Prison Calls to Predict Crime

    The most consequential development is the shift from keyword flagging to predictive analysis. MIT Technology Review reported that Securus Technologies began building AI tools in 2023, using its massive database of recorded calls to train models to detect criminal activity — including one model built on seven years of calls made by inmates in the Texas prison system, with state- and county-specific models in development. Over the year preceding the report, Securus had been piloting the tools to monitor inmate conversations in real time, analyzing phone and video calls, texts, and emails and flagging sections for human agents to review. [3]

    Securus president Kevin Elder described the capability in explicitly predictive terms: “We can point that large language model at an entire treasure trove [of data], to detect and understand when crimes are being thought about or contemplated, so that you’re catching it much earlier in the cycle.” The company said its monitoring has helped disrupt human trafficking and gang activity organized from within prisons, and to identify staff bringing in contraband — though it did not provide MIT Technology Review with any cases specifically uncovered by the new AI models. [3]

    Two features distinguish this from ordinary call monitoring, and both cut in the company’s favor on capability and against it on governance.

    • The training data is the calls themselves. The models are built from years of recorded conversations between incarcerated people and their families. People are told calls are recorded, but there is no meaningful consent to having their voices used as training data. Bianca Tylek of the prison rights group Worth Rises called it “coercive consent; there’s literally no other way you can communicate with your family” — and noted that in most states, “not only are you not compensating them for the use of their data, but you’re actually charging them while collecting their data.” The core logic of the objection: consent cannot be free when the alternative to agreeing is isolation. [3]
    • The target has moved from crimes committed to crimes contemplated. Keyword systems surface evidence of offenses that have occurred; a model tuned to detect “when crimes are being thought about or contemplated” is aimed at intent. That is the move from surveillance to prediction — the same conceptual step that has drawn scrutiny in every other domain where AI has entered criminal justice, from facial recognition to risk assessment. [3]

    Securus has a documented history of capturing privileged conversations. Leaked recording databases showed the company improperly recorded thousands of calls between inmates and their attorneys; it settled a California class action for $880,000 over unlawfully recorded public-defender calls, and it and CoreCivic paid $5.3 million to settle two Kansas suits from some 500 detainees whose attorney calls were recorded. The New York Department of Investigation found Securus inadvertently recorded about 324 privileged calls in a single system in roughly a year — a small subset of about 700,000 calls analyzed, but one that the ACLU’s National Prison Project says shows how few limits courts have placed on this surveillance power. “I think this is one of many situations where the technology is way far ahead of the law,” said Corene Kendrick, the project’s deputy director. [3]


    The FCC Reversal: Making Inmates Fund the Surveillance

    The FCC’s October 28, 2025 order (FCC 25-75) is the policy hinge of the whole story, and it turned on a specific statutory instruction. The Martha Wright-Reed Act of 2022 directed the Commission to consider “costs associated with any safety and security measures necessary to provide” communications services to incarcerated people. In 2024, a unanimous Commission read that mandate narrowly, capping rates and forbidding telecoms from passing recording and security costs to inmates — and required facilities to fund monitoring from their own budgets. The 2025 order reversed course on the central question of who pays. [4]

    Under the new rules effective April 6, 2026, per-minute caps rose across every facility tier — a prison call that would have cost 6 cents per minute under the 2024 reform now caps at 11 cents, and video call caps run to 25 cents per minute in prisons — and providers may add a separate “rate additive” on top. More consequentially, the FCC expanded the cost categories providers may recover from incarcerated users from two to seven, adding Law Enforcement Support Services, Communication Recording Services, Communication Monitoring Services, Voice Biometric Services, and Other Safety and Security Measures. The Communications Recording Services category explicitly includes “storing recorded communications, transcribing such recordings, and converting recordings into digital formats to support investigation and litigation activities” — and the Communications Monitoring Services category is, per the order’s own accounting, now predominantly AI. [4]

    Commissioner Anna Gomez dissented, and her statement is the sharpest available framing of the governance stakes:

    “Law enforcement,” she wrote, “should foot the bill for unrelated security and safety costs, not the families of incarcerated people.” [5]

    Gomez’s dissent identified the evidentiary core of the dispute: by the FCC’s own record, 76% of prison systems already complied with the 2024 caps, with only Florida, Kentucky, and Oklahoma above even the new interim caps — meaning most facilities could and did provide service within the lower rates. The evidence of crisis the order relied on, she argued, amounted to a single Arkansas sheriff’s office that ended phone service, a handful of small jails where providers stopped service, and speculative statements from sheriff associations. The 2024 data collection found security costs were not “used and useful” to providing communications; the 2025 order reversed that finding based on the providers’ own submissions — the same providers, Gomez noted, whose data quality the order elsewhere described as insufficient. “The FCC today decides to reward bad behavior,” she wrote. [5]

    The core logic of the dissent is a principle that extends well beyond prison phones: when a regulated monopoly funds an activity through user fees, the fee-payers have no exit, and cost-shifting is indistinguishable from a tax levied by private companies. Prison telecom is a monopoly by design — the facility picks the provider, and incarcerated people cannot switch. Putting AI monitoring costs on the rate base therefore means the people under surveillance pay for the instruments surveilling them. [4] [5]


    Why It Matters Beyond Prisons

    The prison call story is a compact case study in how AI becomes infrastructure without ever being voted on as AI policy. Consider the pattern:

    • Power first, automation second. Every state already permitted call monitoring; AI vendors simply removed the human-review bottleneck. The same pattern governs most AI in criminal justice — the legal authority pre-exists, and the algorithm changes only the scale. This is why dedicated AI statutes keep arriving after the deployment, not before it. [1]
    • Budget lines are AI policy. Ohio’s $1 million appropriation, Cuyahoga’s $2.7 million contract, and the FCC’s rate-additive all function as AI authorizations without any of them being an AI law. Legislatures and agencies are already making AI policy every budget cycle; they are just not calling it that. [1] [2] [4]
    • The surveillance extends beyond the surveilled. Every monitored call has two parties, and the person outside the prison walls has not been convicted of anything — often has not been charged with anything. Advocates note the systems produce disproportionate false-positive flags on Black and minority incarcerated callers and struggle with non-Eurocentric accents, extending surveillance’s reach into what one advocate called “already over-policed communities of color.” [1]
    • Self-reported efficacy data is the accountability system. Cuyahoga’s 88 confessions and 30 prevented suicides come from the department pitching its own contract extension. No independent audit of the technology’s accuracy, false-positive rate, or disparate impact is required anywhere in the country before deployment. Where such scrutiny has occurred in adjacent domains — facial recognition, risk assessment — it has generally found error rates concentrated on the same populations. [2]

    Where Policy Goes From Here

    Several concrete governance moves are visible in the current landscape, each addressing a different layer of the problem:

    • Privilege protection by technical requirement, not policy. The repeated attorney-call recording failures at Securus and its peers suggest that registration-and-safeguard procedures are not robust at scale. Requiring technical privilege filtering — with vendor liability for privileged captures — would move the protection from administrative discipline to engineering. [1] [3]
    • Data-use limits on training. The coercive-consent problem is specific to AI: recording is disclosed, but model training is not. A notice-and-consent requirement for training use, or a prohibition on training commercial models on privileged or family communications, would address the gap directly. [3]
    • Independent efficacy audit before scale. Cuyahoga’s numbers deserve the same scrutiny any evidence-generating technology gets in court: validated accuracy, measured false-positive rates, and disparate-impact testing, audited by parties who do not hold the contract. The FCC’s 2025 order ordered another mandatory data collection on rates; nothing comparable exists for the monitoring technology itself. [4]
    • Who pays. The Gomez dissent articulates the cleanest principle on offer: if AI monitoring is a law enforcement tool, law enforcement budgets should fund it. The FCC’s further rulemaking on permanent caps is the open venue for revisiting that choice — the Commission is still seeking comment on the new rules before final effect. [5]

    The prison phone system is where three pressure points of AI policy intersect: monopoly infrastructure, an unexitable user population, and a security rationale that makes every incremental expansion politically easy. What happens in this niche is worth watching for a simpler reason: it is the most complete working model of AI surveillance funded by the people it watches.


    References

    [1] Signal Ohio — [SECONDARY] “Ohio prisons to use AI to listen in on ‘all inmate’ calls.” 2025.

    [2] Ideastream Public Media — [SECONDARY] Richmond, Matthew. “Cuyahoga County considers AI contract to monitor jail phone calls.” September 16, 2026.

    [3] MIT Technology Review — [SECONDARY] “An AI model trained on prison phone calls now looks for planned crimes in those calls.” December 1, 2025.

    [4] Federal Communications Commission — [PRIMARY] “Incarcerated People’s Communications Services (formerly Inmate Calling Services).” Updated January 14, 2026.

    [5] FCC — [PRIMARY] Gomez, Anna M. “Statement of Commissioner Anna M. Gomez, Re: Incarcerated People’s Communications Services.” October 28, 2025.


    AI automated and generated content. Human reviewed.

  • The Wide Net: How Eighty Memos, a New Extremism Label, and a 71-Percent Poll Number Put AI Skeptics in the Intelligence Pipeline

    The Wide Net: How Eighty Memos, a New Extremism Label, and a 71-Percent Poll Number Put AI Skeptics in the Intelligence Pipeline

    Three document caches published in 2026 — more than a thousand pages of internal FBI, DHS, and fusion center reporting — reveal a two-track doctrine taking shape inside American law enforcement: generative AI as a terrorist’s planning tool, and opposition to AI infrastructure as a nascent extremist ideology. The second track is colliding with a public that already opposes local data centers by a wide margin, and the documents show the line between threat assessment and opinion monitoring getting harder to find.


    What the Three Caches Show

    On October 14, 2026, the Guardian published the largest look to date at how American law enforcement thinks about artificial intelligence: more than 80 previously unreported documents and memos, distributed among federal, state, and local agencies — including DHS, the FBI, and the NYPD — between 2021 and 2026, obtained by the transparency nonprofit Property of the People.[1] The cache shows agencies pursuing two concerns at once: extremists using generative AI to plan and promote attacks, and what officials describe as a rising threat from people who oppose the technology itself.

    Two earlier investigations reached the same fork by different routes. In May, WIRED reported on more than 1,000 pages of unpublished reports from DHS, the FBI, and regional fusion centers documenting a national shift toward surveilling a new and loosely defined category of “anti-technology extremists.”[2] In June, the Intercept published a confidential December bulletin from the Delaware Valley Intelligence Center, a fusion center housed inside the Philadelphia Police Department, that combed anti-AI social media commentary and warned of a growing physical threat to data centers — while acknowledging it had no specific information about any actual plot.[3]

    Read together, the three caches describe an intelligence enterprise that has simultaneously become dependent on AI tools, fearful of AI-enabled crimes, and protective of the AI infrastructure now embedded in its mission. That triple role is the structural fact underneath the story: when an agency both uses a technology and guards it, criticism of the technology starts to look operationally relevant — and that is where the policy problem begins.


    Track One: AI as the Attacker’s Tool

    The first track is grounded in real cases. In May 2025, a 25-year-old man detonated a car bomb outside a Palm Springs fertility clinic, killing himself and injuring four others; a previously unreported law enforcement document reviewed by the Guardian states the suspect used a generative AI tool that failed to detect his request for bomb-making instructions, and that the private nature of the chatbot exchange prevented any tip-off to police.[1] A May 2026 memo from the FBI’s Los Angeles division predicts the pattern will grow:

    “Threat actors will likely increase their use of GenAI tools for attack planning. Current built-in GenAI safeguards are insufficient to prevent illicit use as malicious actors successfully employ ‘jailbreaking’ techniques to access dangerous information.”[1]

    The documents track this fear across several crime types, and each shows the same underlying logic — AI collapses the distance between having a violent intent and having operational capability:

    • Attack planning. An April 2024 DHS intelligence memo suggested, without citations or evidence, that foreign terrorist organizations and domestic violent extremists had “successfully bypassed generative AI guardrails” to obtain bomb-making instructions and attack tactics.[1] The evidentiary gap matters: agencies were writing policy-relevant warnings before they had confirmed cases.
    • Child sexual abuse material. An April 2025 series from the Joint Regional Intelligence Center, the largest DHS fusion center in the country, flagged a surge in AI-generated CSAM, citing National Center for Missing and Exploited Children data of more than 7,000 child exploitation reports involving generative AI between 2023 and 2025.[1]
    • Propaganda and radicalization. A memo from the Orange County Intelligence Assessment Center warned that individuals and groups were weaponizing AI “to spread hate speech, recruit new members, and radicalize online supports more efficiently” — though one of its cited examples, a January 2025 school shooter, was linked to AI via a source that contained no mention of AI.[1]
    • Evidence fabrication against police. By 2026, intelligence centers were warning that AI-generated fake evidence was being used to harass officers, including a Florida man arrested in March for showing an officer an AI-generated video of his patrol car being broken into. One memo called the practice “a form of law enforcement harassment” that can “elicit an armed law enforcement response.”[1]

    The documents also flag a quieter structural consequence: private AI conversations lack the “tripwires” that open forums have — bystanders who might report a would-be attacker. The FBI has encouraged agencies to build relationships with AI companies to streamline reporting avenues.[1] That is a reasonable investigative instinct, but it institutionalizes a channel between platforms and police whose scope no legislature has defined.


    Track Two: AI Infrastructure as the Target

    The second track also rests on real cases. In October 2021, Seth Pendley was sentenced to ten years in federal prison for plotting to blow up an Amazon Web Services data center in Virginia, having told a confidential source he hoped the attack could “kill off about 70% of the internet” and provoke a reaction against what he perceived as a “dictatorship.”[4] A joint October 2021 threat analysis from DHS, the National Counterterrorism Center, and the FBI circulated the case as evidence that far-right militias saw critical infrastructure as a target for accelerating societal collapse.[1]

    But in the years since, the documents show agencies widening the aperture well beyond confirmed plots. The Guardian’s cache includes a 2025 Philadelphia police intelligence bulletin quoting an account associated with a white supremacist posting “I bring a sort of ‘tannerite and gasoline’ vibe to ai data centers,” and a Michigan demonstration in which three people allegedly tied to a white supremacist group protested what they called a “Jewish data center.”[1] Multiple fusion centers warned of posters on social media platforms and forums vowing to destroy data centers, and far-right forum users discussing killing AI executives and officials who support datacenter projects.[1]

    The Philadelphia bulletin shows how far downstream this monitoring reaches. It flagged posts from an anonymous anti-capitalist blog with a title invoking the “Butlerian Jihad” — a fictional anti-robot movement from the science fiction novel Dune — urging readers to “smash the screens that are brainwashing us into submission” with “ATTACK! ATTACK! ATTACK!”, a post that named no targets and was paired with a cartoonish sketch of a computer struck by arrows. It also cited a Facebook meme reading “I cannot escape the feeling that I am morally obligated to sabotage AI data center infrastructure.”[5] The bulletin’s core warning was this: “Domestic violent extremists (DVEs) are likely interested in targeting artificial intelligence (AI) data centers, posing a physical and cyber threat to infrastructure in the Philadelphia regional area” — issued while conceding “a lack of specific information on plans to target AI data centers in the Philadelphia area.”[5]

    What makes the bulletin policy-relevant is its list of warning indicators, which placed “disruptive First Amendment activity” among the signs of extremist risk and treated social media criticism of data centers as pre-operational behavior.[3] A report from the Northern Virginia Regional Intelligence Center did the same with Suspicious Activity Reporting categories, listing “observation/surveillance” and “photography” as indicators — activities that lawyers note are also how peaceful protesters, journalists, and residents document projects they oppose.[2]


    A Category That Did Not Exist a Year Ago

    The most consequential finding in the WIRED reporting is terminological. A New York Police Department intelligence assessment warned that “the chaotic atmosphere that may result from emergent AI technology in the next five years may fuel large-scale protests that devolve into civil unrest and anti-tech violent extremist activity, especially in large urban areas such as New York City.”[2] As WIRED documented, the term “anti-tech violent extremism” appears in no publicly available DHS or FBI domestic extremism guides; it is a novel grouping that fuses distinct ideologies — far-right accelerationists, anarchists, environmentalists, and doomsday-focused rationalists — into a single monitored category.[2]

    The category did not emerge in a vacuum. It follows National Security Presidential Memorandum 7, signed September 25, 2025, which directs the Justice Department’s Joint Terrorism Task Forces to investigate and disrupt the networks behind political violence, and identifies “anti-Americanism, anti-capitalism, and anti-Christianity” as common threads animating violent conduct.[6] The administration’s May 2026 counterterrorism strategy named left-wing extremists among the nation’s top three counterterrorism priorities.[2] The institutional logic is familiar from prior threat-category expansions: once a directive establishes a broad motivating-ideology framework, field-level intelligence products organize what they observe to fit it.

    Critics point out what that does to error rates. “These intelligence reports are part of a long tradition of agencies identifying protest or even simply having strong opinions as precursors to violence,” Spencer Reynolds, senior counsel at the NAACP Legal Defense Fund, told WIRED. “Suspicious activity reports are incredibly unreliable, often about vague or innocent behavior, issued under permissive standards.”[2] The core logic of the critique is an asymmetry: a fusion center that fails to warn of a real attack faces institutional blame, while one that over-warns faces none — so the rational institutional move is always to widen the net.


    When the Majority Becomes the Indicator

    The defining feature of this surveillance wave is that the monitored sentiment is no longer marginal. A Gallup survey conducted March 2-18, 2026 — the first time Gallup has asked the question — found that seven in ten Americans oppose constructing AI data centers in their local area, with 48 percent strongly opposed; opposition exceeds the historical high for local nuclear plant construction.[7] Distrust of the AI industry is now a mainstream midterm-election issue, which is precisely what distinguishes this category from the fringe movements fusion centers were built to watch.

    The documents show agencies absorbing that mainstream opposition into threat frameworks rather than separating themselves from it. A Nevada Threat Analysis Center report from April 2026 cited a Reno city council hearing on a datacenter moratorium — where residents raised concerns about water consumption, utility costs, pollution, and climate impacts — as part of a trend of “increased negative sentiment towards the facilities.”[1] The hearing was an ordinary exercise of local democracy; in the intelligence product, it became a data point about sentiment trends. Roughly 80 fusion centers, created after September 11 to bridge federal intelligence and local policing, are the pipeline through which these products circulate nationwide.[2]

    Philadelphia civil rights lawyer Paul Hetznecker, responding to the Delaware Valley bulletin, put the objection plainly: “Those are legitimate, popular political concerns that are raised by local communities.” He described the report as “a very dangerous attempt to characterize that protected First Amendment activity — activity which is fundamental to our democracy — as something other, something more dangerous, a breeding ground for something more sinister.”[3] The agencies see it differently. A Philadelphia Police Department spokesperson said the center “recognizes and respects the rights of individuals to lawfully express opinions,” and that its assessments “are designed to provide situational awareness, not to characterize lawful activity or constitutionally protected speech as criminal conduct.”[3] The NYPD said its report merely summarized the “AI 2027” forecast for officer awareness, and DHS said it “remains committed to mitigating all forms of ideologically-motivated violence” while assessing that malicious actors “may seek to use artificial intelligence to enhance existing harmful activities.”[1]


    What the Profession Says It Should Do

    Law enforcement’s own professional bodies have been building the alternative to category-driven monitoring. In July 2026, the Future Policing Institute — a practitioner-focused organization led by former Redlands police chief Jim Bueermann and promoted through the International Association of Chiefs of Police — released its Guiding Principles for Responsible AI in Policing, ten principles covering privacy, algorithmic fairness, transparency, and ongoing review.[8] The framework rests on three commitments that bear directly on the fusion center problem:

    • Human judgment must prevail. Consequential decisions about arrests, force, and deprivation of liberty must belong to trained professionals, not systems — which means an intelligence product can inform discretion but cannot substitute for it.[8]
    • Reasonableness as the baseline. The principles extend the existing Law Enforcement Code of Ethics to new technology rather than inventing a new ethic — the implicit claim being that agencies do not need novel categories like “anti-tech extremism” to judge conduct; they need to apply the standards they already have to the conduct actually observed.[8]
    • Transparency before deployment. Agencies should publish plain-language information about the tools they use and engage communities honestly before deployment, not after.[8]

    The framework’s stated endpoint is worth noting for what it implies about the current practice: “the goal of AI in policing is equal justice, not merely public order.”[8] A monitoring regime that sweeps majority opinion into the extremist file is optimizing for public order by default; the profession’s own standard asks for more.


    The Line-Drawing Problem

    The difficulty is that both tracks in these documents are real. Pendley genuinely tried to bomb a data center; the Palm Springs bomber genuinely used a chatbot; violent posts genuinely exist. The policy failure the caches reveal is not invention of threats but the absence of a firewall between confirmed threat and protected sentiment — the same bulletin that analyzes a genuine bomb plot also catalogs memes and a city council hearing, and nothing in the product distinguishes their evidentiary weight.

    Three design questions follow, and they are the ones legislatures and agencies will eventually have to answer:

    • What counts as an indicator? If “photography,” “attending hearings,” and “disruptive First Amendment activity” are warning signs, the category has no floor. Indicator lists need to be tied to conduct reasonably suggestive of operational planning, not opinion.
    • Who audits the category? “Anti-tech violent extremism” was coined in field products, not statutes. Novel threat categories that drive surveillance should require some external validation — from inspectors general, state legislatures, or the DOJ units that set domestic terrorism policy.
    • What does the public get to know? The fusion center network operates largely through “for official use only” products. The FPI transparency principle suggests the inverse default: agencies monitoring a technology debate should disclose that they are, what they collect, and under what rules.[8]

    The sharpest warning in the documents comes from the monitored side. “There’s nothing extreme about not wanting to lose your livelihood to AI, your drinking water to AI datacenters, your privacy to ubiquitous AI surveillance or your life to runaway rogue AI,” said Ryan Shapiro, executive director of Property of the People. “What’s extreme is law enforcement treating these legitimate objections as threats to be neutralized.”[1] When seven in ten Americans hold the position being monitored, the distinction between protecting infrastructure and policing opinion is no longer a philosophical nicety — it is the whole ballgame.


    References

    [1] [SECONDARY] The Guardian — “Tool and target: US law enforcement grapple with AI-enabled terrorism and anti-tech extremism” by Johana Bhuiyan (October 14, 2026)

    [2] [SECONDARY] WIRED — “US Law Enforcement Warns of ‘Anti-Tech Extremism’ as AI Hatred Grows” by Daniel Boguslaw (May 26, 2026)

    [3] [SECONDARY] The Intercept — “Philly Cops Admit That They’re Tracking ‘First Amendment Activity’ Critical of AI” by Matt Sledge and Sam Biddle (June 1, 2026)

    [4] [PRIMARY] U.S. Department of Justice, Northern District of Texas — “Texas Man Sentenced to 10 Years for Plotting to Attack Data Centers” (October 1, 2021)

    [5] [PRIMARY] Delaware Valley Intelligence Center — “Data Centers Bulletin” (December 2025), via DocumentCloud

    [6] [PRIMARY] The White House — National Security Presidential Memorandum 7, “Countering Domestic Terrorism and Organized Political Violence” (September 25, 2025)

    [7] [PRIMARY] Gallup — “Americans Oppose AI Data Centers in Their Area” (May 13, 2026)

    [8] [ADVOCACY] Future Policing Institute (policing professional organization) — “FPI Develops Guiding Principles for Responsible AI in Policing” (July 7, 2026)


    AI automated and generated content. Human reviewed.

  • The Unfinished Draft: What One Year Under America’s Only Two AI Police-Report Laws Revealed

    The Unfinished Draft: What One Year Under America’s Only Two AI Police-Report Laws Revealed

    Utah and California enacted the nation’s first statutes governing AI-written police reports. Twelve months later, the bills that would have followed them lapsed or died in four legislatures, prosecutors quietly became the field’s real regulators, and the machine’s first draft — the document both laws were built to preserve — remains the system’s missing evidence.


    Two Statutes, Two Theories of the Problem

    Police reports look like paperwork, but they are the narrative the rest of the criminal process inherits: they shape charging decisions, plea negotiations, sentencing, civil liability, and a defendant’s ability to challenge the case.[1] Since 2024, vendors led by Axon’s Draft One have sold tools that generate draft report narratives directly from body-worn camera audio, and as of October 2026 exactly two states have enacted statutes governing the practice — with no binding federal law on the books.[2]

    Utah moved first, with a disclosure theory. SB 180, signed March 25, 2025 and effective May 7, 2025, requires every law enforcement agency to adopt a written generative-AI policy — spelling out which tools employees may use, which tasks they are permitted for, the importance of reviewing AI content, and the acknowledgment that violations can bring administrative discipline. Any report or record created wholly or partly with generative AI must carry a disclaimer, and the author must certify they read and reviewed it for accuracy.[3] The logic is that sunlight is the cheapest, fastest intervention: force the agency to admit on the face of the document that a machine wrote part of it, and judges, defenders, and juries can weigh it accordingly. What Utah does not require is keeping the AI draft, which means no one can later compare what the machine produced against what the officer signed.

    California moved second, with an evidence theory. SB 524, signed in October 2025 and effective January 1, 2026, added Penal Code section 13663. When a report is generated fully or partly by AI, the agency’s policy must ensure:[4]

    • Page-level disclosure. Every AI program used must be identified in a manner “readily apparent to the reader,” alongside the sentence: “This report was written either fully or in part using artificial intelligence.”[4]
    • Officer certification. The preparing officer signs, physically or electronically, verifying they reviewed the report and that the facts are true and correct.[4]
    • First-draft retention. The initial AI draft is retained for as long as the final report — the provision that matters most, because it creates a before-and-after record of human editing.
    • A legal firewall on authorship. Except for the final report, no AI draft “shall not constitute an officer’s statement” — only the signed final version counts, closing the loophole where an officer blames the machine for inaccuracies.[4]
    • An audit trail identifying who used the AI and what video or audio footage the draft was built from.[4]
    • A vendor data ban. Contractors may not share, sell, or reuse agency information except for the agency’s own purposes or pursuant to court order.[4]

    Author Senator Jesse Arreguín (D-Oakland) framed the stakes in liberty terms at the signing:

    “We’re not going to gamble with personal liberty. AI hallucinations happen at significant rates, and what goes in a police report can influence whether or not the state takes away someone’s freedom.”[5]

    Neither law bans anything. Both assume that officer review can carry the constitutional weight. The year’s evidence tested that assumption from three directions — the research, the legislatures, and the prosecutors.


    The Evidence Arrived After the Signings

    The strongest independent study contradicts the efficiency case that drove adoption. A randomized controlled trial by Adams and colleagues, published in the Journal of Experimental Criminology, found that AI did not significantly improve police report-writing speed.[6] That matters because vendor marketing — Axon claims its tool can cut report-writing time by up to 80 percent — is the primary justification agencies give for adoption, and a follow-up survey titled “A Good College Essay but a Bad Police Report” found the AI output was overall less accurate than narratives written by humans.[7]

    The failure modes are not hypothetical. In Heber City, Utah, a department testing AI report software watched a draft absorb a nearby television playing “The Princess and the Frog” into the official incident narrative — an officer transformed into a frog in the record, caught and corrected only because the error was absurd enough to notice.[7] King County, Washington’s “phantom officer” draft became the second documented case, and both now circulate as the field’s shorthand for what unreviewed output looks like.[1] The uncomfortable lesson is that the errors easy to catch are the ridiculous ones; a plausible-sounding hallucination — a wrong sequence of events, an inferred intent — is precisely the kind a tired officer reviewing a polished draft will miss.

    The Council on Criminal Justice’s September 2026 case study, applying its AI decision framework to a hypothetical 200-officer department, classified AI report writing as carrying both substantial risk and substantial opportunity — a “careful implementation” rating, not a green light. It flagged automation bias as an acute danger, since officers may treat a polished draft built from body-camera data as an authoritative account, and recommended that agencies require officers to document independent recall before reading the AI draft, to keep the machine’s narrative from reshaping the officer’s memory.[1] King County’s chief deputy prosecutor made the same point from the other direction: while an officer must edit the narrative and attest to its accuracy, “some of the [genAI] errors are so small that they will be missed in review.”[8]


    Legislatures Stalled. Prosecutors Moved.

    The 2026 sessions were supposed to extend the Utah–California precedent. Instead, the pipeline jammed:[2]

    • New York’s S10425/A9253-A — the most ambitious bill anywhere, defining “Covered AI” to expressly include report-writing tools, with a public inventory, disclosure, first-draft retention, and enforcement by both the attorney general and private plaintiffs — lapsed when the Legislature adjourned sine die on June 5, 2026. Reintroduction is expected in 2027.[2]
    • Minnesota’s SF 4575/HF 4536, which would have prohibited generative AI in official records including police reports outright, died without a hearing at the May 2026 adjournment. Nevada’s SB 199 ban failed in 2025.[2]
    • Utah’s own SB 205, a follow-up that would have required annual policy review, locked AI safeguard settings, and public posting of agency policies, passed the Senate and died in the House in March 2026.[2]
    • Virginia’s HB 1294 was carried over to 2027; New Jersey’s A3076, directing an Attorney General study, remains pending.[2]

    While the legislatures idled, the officials who actually decide what enters a courtroom did not:

    • Connecticut’s Chief State’s Attorney imposed a statewide moratorium on using AI to draft or narrate criminal reports in April 2026 — the broadest restriction in the country, and notably not a statute at all.[2]
    • King County, Washington has refused AI-assisted narratives since a September 2024 memo: “For now, our office has made the decision not to accept any police narratives that were produced with the assistance of AI.”[8]
    • Wayne County, Michigan Prosecutor Kym Worthy wrote to every police chief in the county in September 2026 urging “cautious and restrained” use of AI-generated reports and announced a countywide policy; Hays County, Texas commissioners voted down funding for Axon’s AI tools after a sheriff’s office trial, with the Kyle Police Department declining to use them at the district attorney’s request.[2]

    The core logic here is leverage, not ideology. A department can buy any tool it wants, but if the prosecuting attorney will not file charges on an AI-assisted report, the product has no practical value in that jurisdiction. Prosecutors move by memo, need no supermajority, and answer to no vendor lobby — which is why the real regulatory map of AI police reporting in 2026 is drawn less by statute than by a scatter of county lines and one statewide moratorium.


    The Gap Neither Law Closes

    Beneath the disclosure and retention rules sits a problem both statutes left untouched: the system prompt — the vendor’s hidden instructions that shape every narrative the model produces. The Council on Criminal Justice concluded that prompt opacity creates “a structural accountability gap that is not fully addressed by CA SB 524 or Utah SB 180,” because the prompt materially shapes the report but is not retained, versioned, or produced in discovery.[1] Its recommendation was unusually blunt:

    “Vendor refusal to retain and produce prompts in discovery should be treated as a Generally Avoid trigger, not a negotiating posture.”[1]

    The doctrinal stakes are Brady and the Sixth Amendment. Existing discovery regimes were not designed for an evidentiary record with a prompt, a model version, and an editing history between draft and final report, and officers may be unable to testify clearly about which sections were machine-written, which they edited, and what they independently remember — the very distinction cross-examination exists to test.[1] Kate Chatfield of the California Public Defenders Association put the defense view in one line at the SB 524 signing: “Too often, police narratives become the default truth, even when they contain errors or biases.”[5]

    Utah Senator Stephanie Pitcher, the former prosecutor who sponsored the first law, anticipated the whole arc. Of her bill she said, “I felt like it was something that it’s better for us to get ahead of, instead of playing catch-up on” — and getting ahead meant accepting that a disclosure requirement, without the draft and the prompt behind it, only tells the court that AI was involved, not what the AI actually produced.[7]


    What to Watch

    • The New York reintroduction (2027). S10425’s Covered AI framework — public inventory, first-draft retention, dual AG and private enforcement — is the most complete legislative template on the table; whether it survives a second session will define the statute map’s next move.[2]
    • Proposed Federal Rule of Evidence 707 on machine-generated evidence, now pending, would give federal courts a reliability test that AI-drafted police reports would have to clear.[2]
    • The first prompt-production fight. When a defense team demands the system prompt and draft history under Brady, the answer will either force vendors to retain what they now discard or expose the prompt gap as a constitutional defect.[1]
    • Prosecutor-policy contagion. Wayne County and Hays County showed that a single prosecutor’s letter can end AI report writing in a jurisdiction overnight. The next metro district attorney to follow Connecticut’s statewide lead will redraw the practical map more than any 2027 bill.[2]
    • Vendor compliance with California. Draft One was built to delete its first drafts. SB 524’s retention mandate forces every vendor operating in California to either rebuild for auditability or drop the market — the first real test of whether an evidence-preservation law can beat a product designed for ephemerality.[4]

    The one-year verdict is not that the laws failed. It is that they bought transparency at the front of the pipeline — a label, a signature, a paper trail — while the harder questions moved downstream: whether the review the laws presume actually catches small errors, whether the prompt that shapes every draft will ever see daylight, and whether prosecutors will keep making policy in the space the legislatures left open.


    References

    [1] [SECONDARY] Council on Criminal Justice. “Assessing AI in Action: A Case Study on Police Use of AI-Generated Report Writing Tools.” September 2026. counciloncj.org

    [2] [SECONDARY] “AI Police Report Laws by State: 2026 Tracker.” Barricade AI (vendor-maintained tracker; the company sells AI tools to law enforcement and discloses its commercial interest). October 1, 2026. barricade.tech

    [3] [PRIMARY] Utah State Legislature. “S.B. 180: Law Enforcement Usage of Artificial Intelligence.” 2025 General Session. Signed March 25, 2025; effective May 7, 2025. le.utah.gov

    [4] [PRIMARY] California Penal Code § 13663 (added by S.B. 524, 2025–2026 Regular Session). “Law enforcement agencies: artificial intelligence.” Effective January 1, 2026. codes.findlaw.com

    [5] [PRIMARY] Office of California Senator Jesse Arreguín. “Governor Signs Bill Requiring Law Enforcement to Disclose Use of AI in Police Reports.” October 13, 2025. sd07.senate.ca.gov

    [6] [SECONDARY] Adams, Ian T., Madison Barter, Kristin McLean, Heidi Boehme, and Ian A. Geary. “No Man’s Hand: Artificial Intelligence Does Not Improve Police Report Writing Speed.” Journal of Experimental Criminology 22 (2024): 137–154. doi.org

    [7] [SECONDARY] Seacrest, Logan (R Street Institute). “The Risky Marriage of AI and Police Reports.” Governing (opinion). July 15, 2026. governing.com

    [8] [ADVOCACY] Guariglia, Matthew. “Prosecutors in Washington State Warn Police: Don’t Use Gen AI to Write Reports.” Electronic Frontier Foundation (digital rights organization). October 17, 2024. eff.org


    AI automated and generated content. Human reviewed.

  • The Residue Problem: How a February Ban, a Dozen Silent Case Files, and a Rules-First SOP Are Reshaping Police AI in Wisconsin

    The Residue Problem: How a February Ban, a Dozen Silent Case Files, and a Rules-First SOP Are Reshaping Police AI in Wisconsin

    Milwaukee banned facial recognition in February, yet prosecutions built on the technology continue — while a generative AI policy written before any use, and a county board’s vote to defund Flock cameras, show police AI governance moving from statutes into directives, discovery practice, and budget amendments.


    A Ban That Arrived After the Practice

    On February 6, 2026, Milwaukee Police Chief Jeffrey Norman issued a department directive banning all Milwaukee Police Department members from using or acquiring facial recognition technology, effective immediately[1]. The announcement came the day after residents packed a Fire and Police Commission meeting and testified for nearly three hours against a proposed agreement with the company Biometrica — a deal that would have provided the department with two “free” software licenses in exchange for millions of images and data from Milwaukeeans[2]. Under the directive, MPD will not consider acquiring the technology again until the department develops a formal policy with the Fire and Police Commission[3].

    What made the hearing combustible was the department’s own April 2025 acknowledgment: for two to three years, MPD had used other agencies’ licenses to obtain facial recognition search results without any written department policy governing the practice — and, according to the ACLU of Wisconsin, without the Fire and Police Commission’s knowledge[1]. The ACLU, which had organized much of the nine-month opposition campaign, called the ban welcome but conditional. “This ban can’t just be lip service or a hollow gesture meant to ease public backlash,” said Advocacy Director Amanda Merkwae. “It must actually mark an end to the use of facial recognition by Milwaukee police” (advocacy organization)[2].

    The episode’s core logic is about sequencing. A technology deployed for years without a policy generates cases, habits, and data before any rule exists to govern them; a ban issued afterward cannot reach backward. That asymmetry — rules that arrive after the practice instead of before it — is now the defining feature of police AI governance in Wisconsin, and the two institutions that just learned it are testing opposite corrections: one inside the courtroom, one inside the rulebook.


    The Cases the Ban Did Not Touch

    Seven months later, the directive has a residue problem. Although MPD halted its use of facial recognition in February, cases in which the technology was used at some point in the investigation are still being prosecuted in Milwaukee County[1]. A department can stop using a tool; it cannot un-use it in cases already charged. The moratorium governs future investigations while the pipeline of existing ones runs on — which is precisely where the technology’s documented error rates, and its disclosure record, still live.

    Milwaukee County District Attorney Kent Lovern said his office treats facial recognition as an investigative lead. “It is not, in and of itself, probable cause,” Lovern said. “It’s not, in and of itself, the sole basis for gaining a search warrant”[1]. Information his office receives about facial recognition use in an investigation, he said, is turned over to the defense as part of the case file[4]. Deputy District Attorney Matthew Torbenson described a 2021 home invasion and sexual assault case in which another law enforcement agency’s facial recognition search returned about five potential suspects from surveillance footage; investigators focused on one person based on additional evidence, including an identification by his former spouse, and all five names were disclosed to the defense. Torbenson argued the error risk would be more concerning “if just facial recognition were being used to directly identify a suspect” — which, he said, “is not at all how we’re using it”[1].

    The defense bar’s account is darker, and it turns on a single question: whether anyone told the defendant. The ACLU of Wisconsin reviewed police reports from more than a dozen cases MPD had identified as examples of facial recognition use. “In many of those cases, there was just no mention of (facial recognition technology) use at all,” Merkwae said. Nobody knows the total number of cases in which MPD used the technology — or, of those, how many produced charges in which the use was disclosed in discovery. “I think that’s still a big unknown,” she said[1]. Chief Deputy District Attorney Jeffrey Altenburg acknowledged that leads developed early in an investigation are sometimes not initially shared with prosecutors, but said that once corroborated by standard investigative techniques, “those details are included in the reports that are provided to our office and turned over to the defendant if they are criminally charged”[1].

    Even when disclosure happens, challenging the result is hard. Connor Romenesko, a forensic science practice coordinator with the Wisconsin State Public Defender’s Office, said defense attorneys can obtain the image submitted for a search and some information about the software used, but not what they would need to evaluate the system itself. “The biggest concern I have with these services is that they’re provided by private companies who don’t necessarily provide information about how these systems are trained,” he said — which makes contesting a facial recognition result “extremely difficult”[1]. Public defenders are now trained to spot vague language such as “investigative lead” as a signal that facial recognition may have gone undisclosed[1]. The stakes of that opacity are documented: federal research has found that facial recognition systems can produce higher false-positive rates for some racial and ethnic groups, including Black, Asian, and American Indian people[1].

    The practical lesson: once the acquisition debate is won, the battleground moves to discovery. A ban controls what police may do next; the case file controls what courts will do with what police already did. For defendants charged on pre-ban investigations, the governing document is not the February directive at all — it is whatever the prosecution’s file happens to contain.


    Rules Written Before First Use

    The department’s approach to its next AI technology inverts the facial recognition timeline entirely. In January 2026, MPD released Standard Operating Instruction 682, “Generative Artificial Intelligence” — a detailed governing policy for a technology the department does not currently use[5]. Where facial recognition ran for years with no written rules, the generative AI rules now precede the tool. The contrast is the lesson institutionalized: SOP 682 is what a department writes after discovering what happens when policy comes second.

    The directive, issued January 23, 2026 and approved by Chief Norman, builds its regime on a few load-bearing rules[6]:

    • Draft-only status. AI-generated content “should be considered draft material only and shall be thoroughly reviewed prior to use” — the technology may supplement human work, never substitute for it[6].
    • Disclosure. Members must indicate AI’s contribution in the final work product, so reviewers can distinguish machine-drafted material from human work[6].
    • Human verification before action. “Prior to taking any law enforcement action based on information from an AI system or tool, the information shall be verified by a human person”[6].
    • Data barriers. Unless contractual safeguards exist, members are prohibited from uploading personally identifiable information, criminal justice database information, or HIPAA data into public AI platforms — generic placeholders like “suspect” or “victim” are to be used instead[6].
    • Governance machinery. An AI committee must meet at least bi-annually; audits must test for biased outcomes and ensure training data is “diverse and representative of the Milwaukee population”; an inventory of approved systems is to be made public when appropriate for transparency[6].

    And one prohibition that functions as the policy’s constitutional clause:

    “Members shall not use generative AI systems to rationalize a law enforcement decision, or as the sole basis of research, interpretation, or analysis of the law or facts related to a law enforcement contact or investigation.”[6]

    The immediate use case on the horizon is report writing. Axon, the company that provides MPD’s body cameras, offers a product called Draft One, which generates draft police reports from body-camera audio — and MPD’s Generative Artificial Intelligence Committee has met to evaluate such products, though a department spokesperson said any onboarding “is only in the consideration phase, so have we given it thought? Yes. Are we anywhere close to implementing that (technology)? No”[5].

    Prosecutors and defenders are already war-gaming what AI-drafted reports do to the courtroom record. Chief Deputy District Attorney Altenburg acknowledged generative AI’s risks but said “it can be a helpful tool as long as it’s overseen and regulated by actual humans making actual observations and making sure that what is generated is consistent with what was intended to be.” Deputy District Attorney Torbenson noted that some evidence cannot be generated at all: in a child neglect investigation, what a home looks and smells like, whether there is a bug infestation, whether food is available — “all of that has to come from the impressions of the officers that are on scene”[5]. Merkwae raised the credibility problem: if an officer’s testimony later conflicts with an AI-assisted report, a contradiction that in a human-written report “could land a prosecutor putting a police officer on the Brady list” — the prosecutors’ record of officers with credibility issues — now raises the harder question of whether the error was the machine’s[5]. And Romenesko sketched the failure mode that needs no malice at all: an officer rushed at the end of a shift fails to adequately review an AI-generated report, and a prosecutor uses it to charge someone[5].

    Read together, SOP 682 is less a technology policy than an evidence-integrity document. Every rule — draft-only, disclosure, human verification, the ban on rationalizing decisions — exists to keep a gap from opening between what the machine wrote and what an officer can actually testify to. The policy’s bet is that the courtroom record, not the vendor demo, is where police AI will be judged.


    The Funding Lever: Dane County and the Whack-A-Mole Problem

    Fifty miles west, a different institution pulled a different lever. On April 16, 2026, the Dane County Board of Supervisors approved 2025 RES-403, amending the 2026 county budget to remove $80,000 in funding for the sheriff’s office’s Flock automatic license plate reader subscription and prohibiting further expenditures on the system[7]. The sheriff’s office had granted access to information gathered by its 24 Flock cameras — an AI network that photographs passing cars and tracks their movements — to more than 140 law enforcement agencies, including some in Iowa, Illinois, Indiana, Michigan, Missouri, New York, and Tennessee[7].

    The board’s reasoning was not that deputies misused the system, but that no one could prove nobody else did. “There are well-documented concerns about how this company operates and uses its technology to violate people’s 4th Amendment rights,” said Board Chair Patrick Miles, adding that the county is “open to considering other companies that have stronger safeguards in place.” Supervisor Chad Kemp, the resolution’s author, put the logic plainly: the sheriff’s office “has not been able to affirmatively confirm that the agencies it shares data with, or Flock itself, are not misusing the information collected by cameras in Dane County. Given what we know about how this technology has been abused elsewhere, the responsible course of action is to end funding for this contract”[7]. The board voted 32-1 in April, and the sheriff confirmed he would comply and stop using Flock when the subscription ran out[8].

    The back story explains why the board’s lever was the budget rather than an ordinance: the sheriff originally installed the system without board approval because it was paid for by a $68,750 grant from Axon Enterprise — the same company whose body cameras and Draft One report-writing product sit in Milwaukee’s technology pipeline[8]. When a surveillance vendor funds the grant, procurement oversight never gets a vote. Dane County’s defunding is the corrective: control the money, control the capability.

    Dane County is one domino in a statewide cascade. The city of Verona pulled its three license plate readers from the Flock network in November 2025 after police officers elsewhere in the country accessed Verona’s cameras on behalf of immigration agents; Flock ignored Verona’s takedown demands for months, and the city eventually covered the cameras with black plastic bags. Monona suspended its use of Flock cameras, and Oshkosh rescinded its contract after its police chief said Flock allegedly misrepresented how its data was used[8].

    But privacy advocates warn that each win substitutes rather than retires the capability. The Dane County sheriff’s office is already exploring other vendors to replace Flock[8], and the board has signaled openness to alternatives with “stronger safeguards”[7]. Jon McCray-Jones, a policy analyst with the ACLU of Wisconsin, calls this the core trap: the conversation has been about “how bad Flock is,” but the underlying issues — similar cameras, similar databases, “similar mass, warrantless tracking” — persist regardless of vendor. “You just have a different logo on the contract,” he said (advocacy organization)[8].

    That is the structural insight Wisconsin’s 2026 skirmishes have produced: the unit of police AI governance is not the technology, and increasingly not even the statute — it is the contract and the budget line. A department directive banned a capability in Milwaukee; a budget amendment defunded one in Dane County; neither, by itself, touches the substitution problem. The capability persists as long as the demand does, which is why activists are now pressing local officials for proactive guardrails on entire categories of AI surveillance rather than company-by-company campaigns[8].


    What to Watch

    • The Fire and Police Commission process. Milwaukee’s facial recognition ban holds only until MPD develops a formal policy with the commission — the moratorium’s conversion into a durable rule, or its quiet expiration into a permanent status quo, is the test[1].
    • Discovery in the residue cases. Whether pre-ban facial recognition leads are disclosed in every active Milwaukee County prosecution is now a case-by-case question — and the total count of affected cases remains, in the ACLU’s words, “a big unknown”[1].
    • The Generative AI Committee’s first product decision. If MPD approves a Draft One-style report writer, the draft-only, disclosure, and human-verification rules meet the pressures of real shifts — and the Brady-list question gets its first test[5].
    • Dane County’s replacement vendor. Whether the sheriff’s next ALPR contract requires board approval, and whether it merely swaps logos, will show whether the county’s new oversight is structural or episodic[7].
    • Wisconsin’s fall elections. Milwaukee officials hoping for state-level surveillance controls are watching the governor’s and Legislature’s races; until the state acts, city councils say they must rely on departments voluntarily restricting their own use[8].

    References

    [1] [SECONDARY] Wisconsin Watch / Milwaukee Neighborhood News Service — “Despite police department ban, facial recognition technology lingers in Milwaukee’s criminal justice system” by Devin Blake (September 13, 2026)

    [2] [ADVOCACY] ACLU of Wisconsin — “ACLU of Wisconsin Calls for Accountability After Milwaukee Police Chief Announces Facial Recognition Moratorium” (February 6, 2026)

    [3] [SECONDARY] Fox6 Now (Milwaukee) — “MPD facial recognition technology banned” (February 6, 2026)

    [4] [SECONDARY] Wisconsin Law Journal — “Milwaukee facial recognition cases raise concerns” by Devin Blake of Milwaukee Neighborhood News Service (September 18, 2026)

    [5] [SECONDARY] Wisconsin Watch / Milwaukee Neighborhood News Service — “Milwaukee Police Department considers use of generative AI” by Devin Blake (September 26, 2026)

    [6] [PRIMARY] Milwaukee Police Department — Standard Operating Instruction 682, “Generative Artificial Intelligence,” issued January 23, 2026 (approved by Chief Jeffrey B. Norman)

    [7] [PRIMARY] Dane County Board of Supervisors — “Dane County Board Votes to Withdraw Funds for Flock License Plate Reader Contract,” Resolution 2025 RES-403 (April 16, 2026)

    [8] [SECONDARY] Bolts / Wisconsin Watch — “Flock loses another Wisconsin contract as AI surveillance backlash spreads” (May 2026)


    AI automated and generated content. Human reviewed.

  • The Prompt Gap: How a ‘Generally Avoid’ Clause, a Stalled Study Bill, and a $30,000-a-Month Cold Case Are Moving Police AI Rules Out of Statutes and Into Contracts

    The Prompt Gap: How a “Generally Avoid” Clause, a Stalled Study Bill, and a $30,000-a-Month Cold Case Are Moving Police AI Rules Out of Statutes and Into Contracts

    California and Utah passed the first disclosure laws for AI-written police reports. A September case study from a national task force says the real decisions are being made somewhere no statute reaches: the procurement contract.


    What the First Disclosure Laws Actually Regulate

    When California Governor Gavin Newsom signed Senate Bill 524 on October 10, 2025, California became the first state to require law enforcement to publicly disclose when generative AI helped draft a police report. The law, codified at Penal Code section 13663 and effective January 1, 2026, applies when AI automatically drafts a report narrative from body-worn or vehicle camera recordings or an officer’s dictated report. It requires agencies to maintain a written policy, to identify every AI program used on each page or clearly within the report, to state prominently that “This report was written either fully or in part using artificial intelligence,” and to obtain the signature of the officer or agency member confirming the report was reviewed and its facts are true and correct.[1] Agencies must retain the initial AI draft for as long as they keep the official report, maintain an audit trail identifying who used the AI and what audio or video informed it, and police their vendors’ ability to share, sell, or reuse agency data.[1]

    Utah got there first in a different register. Senate Bill 180, enacted in 2025 and codified at Utah Code section 53-25-902, requires every law enforcement agency in the state to adopt a policy governing employee use of generative AI, and pairs disclosure of AI involvement with an officer’s certification that the final document is accurate.[2] The two laws are the first generation of their kind, and they share a design: both regulate the output — the finished report and what the public is told about it.

    That design choice has a logic behind it. A police report is not an administrative formality; it shapes charging decisions, plea negotiations, sentencing, impeachment at trial, and civil liability. Disclosure regimes make the AI’s role visible to every downstream reader — prosecutor, defense attorney, judge, jury — on the theory that visibility enables challenge. What they cannot do is reach the layer upstream of the report, where the narrative-shaping instructions actually live.


    The Layer the Laws Cannot Reach

    In September 2026, the Council on Criminal Justice’s national Task Force on Artificial Intelligence published a case study walking a hypothetical 200-officer department through whether to adopt AI report-writing tools. Its central finding is that the disclosure laws do not resolve the deepest problem in the stack: the system prompt. The task force notes that Utah’s and California’s statutes “are first-generation legislative responses that require agency policies around AI use and disclosure, but neither fully resolves the prompt issue.”[3]

    The prompt is the instruction set that tells the model what kind of document to produce — what to include, what to summarize, what tone to adopt. It shapes every report the tool generates, yet under current vendor practice it is typically not retained, versioned, or disclosed. The Electronic Frontier Foundation’s investigation of Axon’s Draft One found that the product processes only audio from body cameras, sprinkles the draft with bracketed placeholders for the officer to fill or delete, and deletes the draft when the officer closes the window — meaning the record of what the machine wrote, and what the officer changed, simply vanishes. EFF concluded that “there is no meaningful way to audit Draft One usage, whether you’re a police chief or an independent researcher, because Axon designed it that way.”[4]

    This matters because of what a police report is in court. The Council on Criminal Justice flags unresolved Sixth Amendment concerns — officers may be unable to testify clearly about which sections of a report the AI generated, which they edited, and what they independently remember — and Brady concerns, since the initial draft or the system prompt may contain material information, “particularly if the final report differs from the initial AI-generated version.”[3] A defense attorney who knows AI was involved but cannot see the draft or the prompt cannot test the most consequential question: did the machine’s version of events survive into the report the jury will hear about?

    The case study’s sharpest recommendation is aimed not at legislatures but at the purchasing department:

    “Vendor refusal to retain and produce prompts in discovery should be treated as a Generally Avoid trigger, not a negotiating posture. If the prompt materially shapes the report but cannot be retained, versioned, audited, or produced when legally relevant, then the agency cannot adequately manage the evidentiary and rights-related risks of the system.”[3]

    The underlying logic is leverage. A statehouse cannot compel a private vendor to expose its system prompt, and a disclosure statute regulates the document, not the pipeline that produces it. But a department signing a contract can demand prompt retention, draft retention, model-change notification, and discovery access as terms of sale. The task force accordingly instructs agencies to treat the system prompt, initial AI draft, and final report “as a set, all subject to discovery” — a posture it describes as going “beyond CA SB 524” — with retention aligned to the longest applicable statute of limitations and a requirement that “tools should not erase initial drafts on export.”[3] In other words: where the statute ended, the contract has to begin.


    The Evidence Problem Underneath

    Why does the prompt gap matter so much? Because the empirical record gives reason to doubt the tools do what vendors say they do. The strongest independent evidence cited by the task force is a randomized controlled trial by criminologist Ian T. Adams and colleagues, published in the Journal of Experimental Criminology, which found that AI did not improve police report writing speed — the core selling point — contradicting vendor time-savings claims.[5] A second study asked experienced police supervisors to review 80 reports, 20 of which had been drafted with Draft One. The 92 reviewers could not identify which reports were AI-assisted at better than coin-flip accuracy, and while the AI drafts scored marginally better on surface measures like grammar, they were rated worse on the measure that decides cases: accuracy, at p = .038, moving a report from roughly the 50th to the 36th percentile.[6] As the ACLU summarized the pattern, “there is not a single independent evaluation that supports the vendor claims being made to police departments right now.”[7]

    The failure modes are not hypothetical. In Heber City, Utah, a Draft One report absorbed the audio of a Disney movie playing in the background during a call and produced a report claiming the officer “shape-shifted into a frog.” “The body cam software and the AI report writing software picked up on the movie that was playing in the background, which happened to be ‘The Princess and the Frog,’” Sergeant Keel of the Heber City Police Department told FOX 13. “That’s when we learned the importance of correcting these AI-generated reports.”[8] A comic error is easy to catch; a confident misstatement of a suspect’s words or an officer’s observations is not. That is the animating worry behind the accuracy findings — an AI draft is not more likely to be obviously wrong, it is more likely to be plausibly wrong, in a document whose polished narrative voice reads as authoritative.

    Downstream actors have already reacted. The King County, Washington prosecutor’s office told police in September 2024 not to use AI to draft official reports, citing error concerns — an early signal that prosecutorial acceptance, not legislation, would be the binding constraint.[9] And some departments have found their own simplest compliance path: the San Diego Police Department issued a training order in December 2025 informing all employees that they “shall not use any AI programs” to author reports unless specifically approved, stating that “the Department does not authorize the use of any generative Artificial Intelligence programs to author a police report” — a ban that doubles as its SB 524 policy.[10]


    Where the Rules Are Actually Being Written

    Between the disclosure statutes and the courtroom sits a large, mostly unregulated middle: the state where departments are buying these tools faster than anyone is writing rules for them. New Jersey is the current case in point. Officers in towns from Mount Holly to Eastampton use AI to draft incident reports, but the state has no statewide strategy or regulation for police AI. A bill to study AI in policing has been introduced twice since May 2025 without advancing, and the Attorney General’s office says only that it “is actively pursuing statewide governance to ensure ethical, constitutional, and fair evaluation and implementation of AI by New Jersey law enforcement,” declining to offer a timeline.[11]

    Some chiefs are not waiting, and their hesitation is itself a form of policy. Middle Township Police Chief Tracey Super told the New Jersey Monitor his department tried an AI tool for a cold case — the 1991 disappearance of 11-year-old Mark Himebaugh — under a free yearlong trial with the company Terawe. “It really didn’t do what we were hoping,” he said. “It was more like a search engine with a database, which was helpful, but nothing that we were able to actually solve Mark’s disappearance.” Continuing would have cost $30,000 a month for a 55-officer department serving about 20,000 residents. “We still want to find Mark, but we can’t devote one tenth of our budget to it.”[11] His larger caution is about contracts as governance: “If you sign a two-year contract with a company, and then you’re six months in and the attorney general says you can’t use that? We’re still obligated to pay that contract.”[11]

    That quote names the actual mechanism. When no statute binds the vendor, the contract is the only instrument that does — and the Council on Criminal Justice’s case study functions as a checklist for using it. Among the terms the task force says a department should require before signing:

    • Prompt and draft retention as contract requirements — prompt retention “beyond CA SB 524,” initial-draft retention beyond the statutory minimum, and mandatory notification when the vendor changes the system prompt or migrates the underlying model.[3] The logic: the audit trail the law requires is worthless if the vendor’s software destroys the underlying records by design.
    • A contractually fixed hallucination-rate methodology, with monthly vendor reporting — “frame as ‘validation, not audit of trade secrets’”.[3] The logic: without a fixed definition of error, the vendor’s performance claims cannot be tested, only repeated.
    • Independent third-party validation as a renewal condition, plus a formal alternatives analysis comparing the AI tool against templates, plain dictation, added clerical staff, and workflow redesign.[3] The logic: the question is never “is AI useful” but “is AI better than the cheaper, risk-free alternative” — a question vendors have no incentive to pose.
    • No “approve all” shortcuts — officer review and edits captured at the per-section level, so the record shows exactly what the machine wrote and exactly what the officer changed.[3] The logic: this is the fix for EFF’s core finding, and it can only exist as a product-design demand at signing time.
    • Indemnification for hallucination-driven evidentiary issues, and termination rights for rights concerns — with the case study predicting vendor pushback precisely there.[3] The logic: liability allocation is where the vendor’s confidence in its own accuracy claims gets priced.
    • A minimum six-month pilot with pre-set success criteria — time savings, error rate, officer edit rate, prosecutor acceptance, defense motions, and demographic parity in output language — and a formal go/no-go decision.[3] The logic: because the harms surface months later in courtrooms, a few weeks of department testing proves nothing.

    The task force classifies AI report writing as both substantial risk and substantial opportunity — its framework’s “careful implementation” category — and ties that classification to annual reassessment, with new triggers including prompt changes, model migrations, bias findings, defense motions, and Brady disclosures.[3] It also prescribes an operational safeguard no statute has attempted: where feasible, officers should document their own independent recollection before reading the AI draft, to keep the machine’s narrative from overwriting the officer’s memory.[3]


    What to Watch

    The pattern to track in the coming year is not another disclosure bill — it is whether the contract terms migrate from a task-force report into actual vendor agreements, and whether “Generally Avoid” becomes what “Generally Avoid” is meant to be: a standard a police chief, a city attorney, or eventually a judge can point to. Two developments would signal that it is taking. First, a public records request surfacing an Axon or competitor contract that contains — or conspicuously lacks — the prompt-retention clause. Second, a defense motion built on the gap itself: a demand for the initial draft and system prompt in a case where the final report differs from what the machine wrote. When that motion is granted, the contract era of police AI governance will have begun in earnest; when it is denied, the prompt gap will have survived its first court test.


    References

    [1] California State Legislature — [PRIMARY] S.B. 524, Law enforcement agencies: Artificial Intelligence (Pen. Code § 13663), 2025–2026 Regular Session. Signed October 10, 2025; effective January 1, 2026.

    [2] Utah State Legislature — [PRIMARY] Utah Code § 53-25-902 (S.B. 180, Law Enforcement Usage of Artificial Intelligence), 2025 General Session.

    [3] Council on Criminal Justice, Task Force on Artificial Intelligence — [SECONDARY] “Assessing AI in Action: A Case Study on Police Use of AI-Generated Report Writing Tools.” September 2026.

    [4] Electronic Frontier Foundation — [ADVOCACY] Guariglia, Matthew, and Dave Maass. “Axon’s Draft One Is Designed to Defy Transparency.” July 10, 2025.

    [5] Adams, Ian T., Barter, M., McLean, K., Boehme, H., and Geary, I. A. — [SECONDARY] “No Man’s Hand: Artificial Intelligence Does Not Improve Police Report Writing Speed.” Journal of Experimental Criminology 22, 137–154. 2024.

    [6] Criminology working paper (crimRxiv) — [SECONDARY] Expert-blind review study of AI-assisted police report quality (80 reports, 20 AI-drafted, 92 reviewers), 2026.

    [7] American Civil Liberties Union — [ADVOCACY] Stanley, Jay. “Studies Question Value of AI-Assisted Police Reports.” May 19, 2026.

    [8] FOX 13 Salt Lake City — [SECONDARY] Constantino, Mya. “Ribbit Ribbit! Artificial Intelligence Programs Used by Heber City Police Claim Officer Turned Into a Frog.” December 19, 2025.

    [9] KOMO News — [SECONDARY] Harris, J. “AI-Assisted Police Reports Not Welcome in King County Due to Potential Errors.” September 2024.

    [10] CBS 8 San Diego — [SECONDARY] “San Diego Police Ban Generative AI in Reports as New State Law Takes Effect.” Training order issued December 2025.

    [11] New Jersey Monitor — [SECONDARY] DiFilippo, Dana. “As Police Increasingly Embrace AI, Some Wait for Regulations and Proof It Works.” September 17, 2026.


    AI automated and generated content. Human reviewed.

  • The Hundred Percent Problem: How a Casino’s False Match, 168 Redacted Names, and a 91 Percent Collapse Are Forcing Police to Corroborate Facial Recognition

    The Hundred Percent Problem: How a Casino’s False Match, 168 Redacted Names, and a 91 Percent Collapse Are Forcing Police to Corroborate Facial Recognition

    Courts, not legislatures, are writing America’s operating rules for police facial recognition. In Reno, a “100 percent match” cost a UPS driver eleven hours in custody and pulled an entire city into federal court. In Detroit, one settlement cut police facial recognition use by 91 percent. The through-line is a single broken assumption: treating a probabilistic similarity score as a positive identification.


    A 100 Percent Match, Eleven Hours in Custody

    On September 17, 2023, the Peppermill Casino in Reno telephoned the police to report that a trespasser had unlawfully returned to the property. The casino’s AI facial recognition software, security officers told responding Officer R. Jager, had provided a “100 percent match” confirming that the man in their security office was M.E. — a patron banned months earlier for sleeping on the premises [1].

    The software was wrong. The man was Jason Killinger, a UPS driver of nearly twenty years, and the identification he was carrying said so in nine places: a Nevada Real ID driver’s license, a Peppermill player’s card, a vehicle registration and insurance card, a Teamsters Local 533 union card, a Department of Transportation medical card, a debit card, and a pay stub — all in his own name [1]. According to the federal complaint, Jager declined to look at any of it, told his supervisor he suspected Killinger had a “hookup” at the DMV manufacturing false IDs, and arrested him so that jail fingerprinting could settle the question [1]. The two men also differed in height and build — the banned patron stood five-foot-nine and 235 pounds against Killinger’s six-foot-one and 200 — a physical mismatch Killinger’s lawyers say should have been obvious on inspection [2].

    Fingerprints at the Washoe County jail confirmed within hours that the man in custody was Jason Killinger and had never been banned from anything. He had nonetheless spent roughly eleven hours in custody, complained of handcuff pain twenty times on body-camera audio, and later received medical treatment for a shoulder injury, the complaint alleges [1], [2]. Prosecutors then pursued a trespass charge under the name “John Doe” anyway, and after a judge dismissed the case the city attorney left the door open to a future identity-theft charge, even offering — and then declining — a referral to a fraud detective [5].


    No Policy, No Training, No Immunity

    Killinger sued in July 2025. In March 2026, U.S. District Judge Miranda Du allowed the case to clear its most consequential procedural hurdle: the City of Reno itself was added as a defendant, exposing the municipality to financial liability [3]. The individual officer kept his qualified immunity for the arrest itself, but the judge drew the line that matters for every police department watching the case. “A municipality is not entitled to assert the defense of qualified immunity,” she wrote [3]. Under Monell v. Department of Social Services doctrine, a city answers for its own policies and customs — and that is precisely what the litigation went looking for.

    What it found is the heart of the case. At his January 22, 2026 deposition, Jager took full responsibility for the wrongful arrest, testifying that it never should have happened — an arrest based on AI facial recognition software, he acknowledged, required corroboration and there was none in Killinger’s case [4]. City attorneys’ own filings described the casino’s software — whose vendor has never been publicly named — as so reliable that Reno officers would request Peppermill’s help identifying suspects with it [4]. And as of April 2, 2026, the amended complaint alleges, the Reno Police Department still had no policy on facial recognition and required no training that a match alone does not establish probable cause [5].

    “Jager’s conduct was not a sporadic incident involving the wrongful actions of a rogue employee, but the result of a widespread custom and practice involving hundreds of municipal employees making thousands of arrests in the same manner over a period of years.” — amended complaint, Killinger v. City of Reno [5]

    The legal core logic here is worth stating plainly. A single officer’s mistake dies with qualified immunity. A custom — a practice so routine that it functions as policy — transfers liability to the city’s treasury. That is why Killinger’s attorney, Terri Keyser-Cooper, called the failure to train officers “when such arrests were being made on a regular basis” outrageous [4]: the allegation is not that Reno employed one careless rookie, but that Reno ran an assembly line on which this arrest was one unremarkable stop.


    The 168 Redacted Names

    If the case becomes a precedent, it will be because of a discovery fight now before Judge Du. In April 2026, Killinger’s lawyers asked for the arrest records of 168 people who were cited or transported to the Reno Police Department following facial recognition identifications at the Peppermill [6]. The city produced more than 1,000 pages of records — with every name and personal identifier redacted, and without first seeking a protective order from the court [6].

    The stakes of the dispute are asymmetrical. Killinger’s side argues the 168 are percipient witnesses whose experiences could establish a pattern of facial-recognition-based arrests, and relevant to the injunctive relief and potential class claims he is pursuing [6]. The city counters that the individuals’ privacy is at stake, that the casino provided only dates, times, and initials, and that only 16 of the 168 arrests occurred before Killinger’s — meaning later arrests cannot show the city had notice of a problem at the time of his [6].

    The deeper significance is scale. Wrongful-arrest litigation has so far proceeded case by case: one plaintiff, one arrest, one settlement. The 168 records would turn one case into a census of a single biometric system’s role in a single city’s arrests — an accounting no American police department has published anywhere. If the names are unsealed, both sides will finally be counting the same thing: how many detentions rested on nothing but the match.


    Detroit’s Counterweight: What One Settlement Bought

    Detroit is the controlled experiment in what happens when a department is forced to change. Robert Williams was arrested in his driveway in front of his family in January 2020 after the department’s system flagged his driver’s license photo in a shoplifting case; his June 2024 settlement, negotiated by the ACLU and the University of Michigan Civil Rights Litigation Initiative, produced what civil rights advocates describe as the strongest police facial recognition policy in the United States [7].

    The settlement’s terms all target the same failure:

    • No arrest on a match alone. Officers are prohibited from making arrests based solely on facial recognition results — the exact conduct Reno’s deposition testimony admitted to [7].
    • No algorithm-derived lineups without corroboration. A photo lineup generated directly from a facial recognition search requires independent evidence linking the identified person to the crime, cutting off the practice of laundering a match through an eyewitness [7].
    • Mandatory training on documented error rates, including the technology’s higher misidentification rates for people of color [7].
    • A retrospective audit of every case since 2017 in which facial recognition was used to obtain an arrest warrant [7].

    The numbers that followed are the clearest natural experiment in American policing on this question. Detroit ran roughly 100 facial recognition searches in 2023, 28 in 2024, and nine in 2025 — a 91 percent decline — and of those nine searches, only one produced an actionable investigative lead [7]. The core logic: when a match can no longer substitute for evidence, officers stop requesting matches. Use did not collapse because the technology got worse; it collapsed toward the cases where the tool could actually carry weight — violent-crime investigations with other evidence to corroborate the lead. A rule that sounds like a procedural nicety turned out to be a volume dial.


    Bloomington: The Reform Template Travels

    On August 18, 2026, a second city signed on to the same approach. Bloomington, Minnesota settled the lawsuit brought by Kylese Perryman, who was arrested in September 2021 after investigators compared surveillance photos of a robbery and carjacking suspect against his booking photo [8]. Perryman was verifiably elsewhere, and his height, weight, and visible tattoos distinguished him from the suspect — yet he spent five days in jail and thirty days on electronic home monitoring, missing two funerals and time with his young family [8]. The suit alleged police never ran a lineup, never contacted eyewitnesses, ignored his alibi, and kept the charges alive even after his attorney produced clear evidence of innocence [8].

    “Bloomington police arrested and jailed an innocent man for days, forcing Perryman to prove his innocence before getting the charges against him dropped,” said Teresa Nelson, ACLU-MN legal director [8]. Perryman himself put it more simply: “I brought this lawsuit because nobody should go through the ordeal that I went through” [8].

    The settlement’s policy changes look small and technical, which is the point. Bloomington agreed to adopt a procedure incorporating the KOPS Alerts guidelines from the state’s criminal justice information system operating manual, and to adopt a procedure stating that officers are prohibited from issuing KOPS Alerts as arrest warrants [8]. A KOPS Alert is an investigative lead broadcast to other agencies — a signal to look further, not a finding of identity. Codifying the distinction between a lead and a warrant moves the line from officer judgment into written procedure, where a violation is no longer a bad guess but a rule breach. The city also paid monetary damages [9], [8].


    Fourteen Arrests, One Broken Assumption

    The pattern behind all three cases is now documented nationally. In April 2026, the ACLU counted 14 publicly documented wrongful arrests caused by police reliance on facial recognition, spanning February 2019 through August 2025, and called for binding restrictions rather than voluntary department guidelines [10]. Nearly every person wrongfully arrested has been Black, though Killinger is white — a fact his own complaint uses to argue the problem is systemic rather than demographic [4]. The demographic skew tracks the technology’s documented error structure: NIST testing of facial recognition algorithms found false-positive rates highest for East and West African and East Asian faces, “with a factor of 100 more false positives between countries” [12].

    The true number is unknowable for a reason that is itself a policy failure. Police departments routinely do not disclose when facial recognition drove an arrest — in at least one documented case, the arrest warrant listed the identification source simply as “a credible source,” with no mention of the software that produced the match [11]. Fourteen is the count of cases that became public through court filings, press coverage, or lawsuits; cases settled quietly, pled out, or never attributed to the technology do not appear in it [10].

    And the problem is scaling up faster than the rules. Law professor Andrew Guthrie Ferguson, author of Your Data Will Be Used Against You, warns that agentic AI is about to invert the investigative process itself: “All that data is going to be dumped into an AI model, and they’re going to query it to say who’s the most likely suspect,” he told Stateline, “and then police and prosecutors have to kind of work backwards to see if it’s accurate” [13]. That flips the traditional sequence — evidence first, suspect second — into its mirror image, and it makes the corroboration question more urgent, not less [13].


    The Regulator’s Ledger: Britain’s 107 Recommendations

    While American rules emerge case by case, Britain’s data regulator is attempting to build the governance layer in advance. Between June 2025 and March 2026, the Information Commissioner’s Office audited five police forces across England and Wales, publishing its findings on August 18, 2026 with 107 recommendations covering compliance and best practice — every one accepted or partially accepted [14]. LexisNexis counts 16 of those as urgent and 54 as high priority, spanning governance, lawful basis, data protection impact assessments, data minimisation, security, accuracy and bias, and information rights [15].

    The audits found forces generally documented a lawful basis for processing biometric data, with compliance notably stronger for live deployments than for retrospective image searches [14]. The urgent items read like a checklist of everything Reno lacked: clear senior oversight, accountability and training for staff, records of what personal information is used and where it comes from, controls over image sourcing and retention, and verification that systems are accurate with steps taken to reduce bias [14]. A sixth audit, of the Metropolitan Police, is scheduled for later this year [14].

    The contrast is instructive. The ICO treats facial recognition primarily as a data-governance problem — documentation, retention, oversight — while American reform treats it as a probable-cause problem. Both are attacking the same failure: a match that enters the system as an answer, with no record of what it actually was. A jurisdiction that requires both — documentation of the match and corroboration before arrest — is doing the only thing that has demonstrably worked: refusing to let the score close the case by itself.


    What the Litigation Ledger Means

    Read together, Reno, Detroit, and Bloomington describe a regulatory vacuum being filled by the only institution that cannot decline the case. Congress has passed no federal rules governing police facial recognition; most jurisdictions impose no binding corroboration requirement [10], [13]. Into that vacuum, each lawsuit deposits one enforced rule:

    • Detroit: a match is not an arrest, a lineup, or a warrant — and the department must audit its own past [7].
    • Bloomington: an inter-agency alert is a lead, not an arrest warrant [8].
    • Reno: a city that runs arrests on uncorroborated matches with no policy and no training answers for the pattern in dollars [3], [4].

    The underlying principle is not anti-technology. Detroit still uses facial recognition — nine times last year, once productively — and Bloomington still broadcasts KOPS Alerts. The rule each case enforced is about epistemics, not equipment: a similarity score is a question, and somewhere between the score and the cell, a human being must supply the answer. Departments that skip that step are not using AI as a tool; they are using it as a verdict.

    What to watch

    • Judge Du’s ruling on the 168 names — whether the largest single-city accounting of facial-recognition arrests proceeds [6].
    • The Killinger trial — a November 2025 settlement conference failed, so the custom-and-practice claims against Reno are headed toward trial [1].
    • Detroit’s retrospective audit of every 2017-to-2023 warrant case that rested on a match [7].
    • The Metropolitan Police audit, which will test whether Britain’s governance framework scales to the country’s largest force [14].
    • Any legislature that converts the Detroit standard — no arrest on a match alone — from a settlement term into a statute [10].

    References

    [1] [PRIMARY] Killinger v. Jager, No. 3:25-cv-00388-MMD-CSD, Complaint (U.S. District Court, District of Nevada, filed July 30, 2025). https://thecivilrightslawyer.com/wp-content/uploads/2026/03/KillingerFederalComplaint.pdf

    [2] [SECONDARY] “Reno Casino AI Misidentification Case Seeks 168 Names.” Hoodline. August 2026. https://hoodline.com/2026/08/reno-man-jailed-11-hours-after-ai-misfire-now-wants-names-of-168-others/

    [3] [SECONDARY] “Lawsuit: Reno police made thousands of unlawful arrests using face ID, suit says.” Reno Gazette Journal. April 7, 2026. https://www.rgj.com/story/news/2026/04/07/facial-id-lawsuit-reno-police-unlawful-arrests/89491408007/

    [4] [SECONDARY] “Judge’s ruling exposes city of Reno to liability in facial ID lawsuit.” Reno Gazette Journal. March 31, 2026. https://www.rgj.com/story/news/2026/03/31/reno-facial-id-lawsuit-rpd/89411008007/

    [5] [SECONDARY] “Lawsuit alleges systemic misuse of facial recognition by Reno police.” Biometric Update. April 2026. https://www.biometricupdate.com/202604/lawsuit-alleges-systemic-misuse-of-facial-recognition-by-reno-police

    [6] [SECONDARY] “Wrongly Arrested Peppermill Guest Demands Identities of 168 Others Flagged by Facial Recognition.” Casino.org. August 2026. https://www.casino.org/news/wrongly-arrested-peppermill-casino-guest-seeks-names-of-168-others-flagged-by-facial-recognition/

    [7] [SECONDARY] Nassar-Smith, Ali. “Detroit Police Facial Recognition Use Drops 91 Percent After Settlement Tightens Policy.” ID Tech Wire. April 27, 2026. https://idtechwire.com/detroit-police-facial-recognition-use-drops-91-percent-after-settlement-tightens-policy

    [8] [ADVOCACY] ACLU of Minnesota. “The City of Bloomington Agrees to Settle Lawsuit After Incorrectly Identifying, Arresting Kylese Perryman.” August 18, 2026. https://www.aclu-mn.org/press-releases/the-city-of-bloomington-agrees-to-settle-lawsuit-after-incorrectly-identifying-arresting-kylese-perryman/

    [9] [SECONDARY] “Bloomington settles lawsuit with man wrongfully arrested due to facial recognition technology.” CBS News Minnesota. August 2026. https://www.cbsnews.com/minnesota/news/bloomington-lawsuit-man-wrongfully-facial-recognition-technology/

    [10] [SECONDARY] “ACLU Says 14 Wrongful Arrests Now Tied to Police Facial Recognition.” ID Tech Wire. April 20, 2026. https://idtechwire.com/aclu-says-14-wrongful-arrests-now-tied-to-police-facial-recognition

    [11] [SECONDARY] “Flawed Facial Recognition Technology Leads Wrongful Arrest and Historic Settlement.” Quadrangle, University of Michigan Law School. Winter 2024-2025. https://quadrangle.michigan.law.umich.edu/issues/winter-2024-2025/flawed-facial-recognition-technology-leads-wrongful-arrest-and-historic

    [12] [SECONDARY] “Arrested by Algorithm: 13 Cases Dismissed, Nearly Every Victim Black.” State of Surveillance. 2026. https://stateofsurveillance.org/news/facial-recognition-wrongful-arrests-13-dismissals-racial-disparity-2026

    [13] [SECONDARY] “Police use of artificial intelligence grows as rules lag behind.” Stateline. June 26, 2026. https://stateline.org/2026/06/26/police-use-of-artificial-intelligence-grows-as-rules-lag-behind

    [14] [PRIMARY] Information Commissioner’s Office (UK). “Facial recognition in policing: earning public trust through strong data protection governance.” August 2026. http://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2026/08/facial-recognition-in-policing

    [15] [SECONDARY] “ICO publishes findings on police use of facial recognition technology.” LexisNexis. August 19, 2026. https://www.lexisnexis.co.uk/legal/news/ico-publishes-findings-on-police-use-of-facial-recognition-technology


    AI automated and generated content. Human reviewed.

  • The Human-in-the-Loop Illusion: How 600,000 AI Police Reports, a Blinded Quality Study, and a Statewide Moratorium Expose the Gap Between Vendor Claims and the Criminal Justice Record

    The Human-in-the-Loop Illusion: How 600,000 AI Police Reports, a Blinded Quality Study, and a Statewide Moratorium Are Separating Vendor Claims From the Criminal Justice Record

    Axon says its AI has drafted hundreds of thousands of police reports and saved 300,000 officer hours. The first controlled studies, a public-records investigation, two state disclosure laws, and a pair of prosecutor moratoria tell a more complicated story about what happens when machines draft the documents that can take away someone’s freedom.


    The Scale of the Experiment

    Two years after Axon launched Draft One, the product that converts body-camera audio into a first-draft police report narrative, the company says the tool is in use by roughly 600 police departments and has helped write more than 600,000 reports, saving an estimated 300,000 hours according to its customer surveys. CEO Rick Smith told shareholders that “AI is the greatest accelerant Axon has ever had,” and the company reported booking approximately $750 million in AI Era contract value in the year after launch, with AI product revenue growing more than 700 percent year-over-year in its most recent quarterly results.[1]

    But a July 2026 Forbes investigation built on public-records requests found departments discovering the opposite of the marketing promise. In Lafayette, Indiana, officers testing Axon’s Form One tool — which fills administrative details like names and license plates rather than the narrative — wrote in internal emails obtained by Forbes: “I know it doesn’t save me time and I know it has inaccuracies that I will have to edit,” and “Form One dramatically increases the time it takes to finish reports.” One officer reported that a form that took 30 seconds to complete manually now took three minutes because of the corrections required.[1]

    The stakes of this dispute are not measured in paperwork hours. Police reports are the raw material of the criminal process: prosecutors use them to decide whether to file charges, judges rely on them at bail hearings, and defense attorneys cross-examine officers against them. Officers spend as much as 15 hours a week writing reports, per Axon’s own data, which is precisely why the product exists — and why the question of whether the drafts are accurate has become one of the most consequential practical AI policy problems in the justice system.[1]


    What the Controlled Studies Actually Found

    The only pre-registered randomized controlled trial of AI-assisted report writing, conducted with a medium-sized police department and published in the Journal of Experimental Criminology, found that AI assistance “did not significantly affect the duration of writing police reports.” The researchers — led by Ian Adams, a former police officer turned University of South Carolina criminologist — tracked 85 officers writing 755 reports, and confirmed the null result with a difference-in-differences analysis covering 6,084 reports over a full year. Officers spent so much time removing irrelevant material, fixing inaccuracies, and adding facts the model missed that the time savings evaporated.[2]

    The study department — Manchester, New Hampshire, the first agency to test Draft One — ultimately stopped using it in 2024, as did the Anchorage Police Department in Alaska, which cited zero time savings. Lieutenant Matthew Barter of Manchester PD, who participated in the research, put it bluntly: “It was just easier to type the report themselves.” Yet when surveyed, about half of Manchester’s officers believed the tool had made them faster. Adams described the result as “a definite gap between the perception and objective reality” — a finding that matters for policy, because officers who believe the tool helps will lean on it more.[1]

    A follow-up study published in May 2026 went after the quality question directly. In a triple-blind experiment, 92 expert law-enforcement reviewers — supervisors with an average of 21.8 years of experience approving reports — evaluated 80 reports from the randomized trial without knowing which had been AI-assisted. The findings cut against the vendor narrative on three fronts:[3]

    • Accuracy was measurably worse. Reviewers rated AI-assisted reports significantly lower on accuracy (p = 0.038), with all six quality dimensions trending negative. The core logic: the tools build narratives from transcripts alone, so they systematically miss what officers see — a bruise on a victim’s arm, the layout of a scene, the tone behind a phrase — and the missing context does not survive into the final document.[3]
    • Detection was impossible. Experts could not distinguish AI-assisted reports from conventional ones, with detection accuracy equal to a coin flip (AUC = 0.50). This is the finding that breaks the oversight model: the entire regulatory architecture being built assumes a reviewing human can catch machine text, and the supervisors whose job it is could not.[3]
    • Approval was unchanged anyway. First-attempt approval rates were identical across conditions, around 22 percent — supervisors approved flawed AI drafts at the same rate as flawed human drafts, meaning the review layer absorbs authorship information it cannot perceive.[3]

    Computational text analysis in the same study found AI reports were more complex, less readable, and written at a higher grade level — the title of the paper, “A Good College Essay but a Bad Police Report,” captures the mismatch. A report read by prosecutors, juries, and judges needs plainness and completeness, and the model optimizes for neither.[3]

    Axon disputes the generalizability of both studies. Spokesperson Alex Engel said the Manchester research offers “an interesting historical snapshot” whose “conclusions rely on AI-assisted narratives generated in 2024 from a single agency,” adding that “AI capabilities and Draft One have evolved significantly in the last two years, and evaluating the quality of an older version of the product should not be generalized to how Draft One operates today across agencies.” No peer-reviewed study has yet tested the newer versions.[1]


    Why the Machine Gets It Wrong

    The Center for Democracy and Technology’s analysis of the tool category identifies a failure chain that starts before the language model ever runs. Body-camera audio is noisy, accented, and full of overlapping speech; even specialized speech-recognition models exhibit word error rates around 10 percent in ideal conditions — roughly one in ten words. Those transcription errors then compound when fed into a generative model, which can misread context, misattribute statements to the wrong speaker, or hallucinate details no one said.[4]

    The most famous failure is also the most instructive. In Heber City, Utah, Draft One absorbed audio from a television playing “The Princess and the Frog” in the background of a recording and merged the animated fairy tale into the official narrative — producing a report stating the officer had transformed into a frog. The reviewing officer caught it before filing. The reason this anecdote matters is not the comedy; it is that a background movie is easy to spot, while the same mechanism — the model importing plausible-but-unverified content into a factual record — produces subtler errors that are not.[4]

    Two structural risks compound the technical ones:[4]

    • Cognitive offloading and automation bias. Officers may anchor on the AI draft rather than their own recollection, accepting text that does not reflect what they witnessed. CDT notes that marketing materials for several vendors promote exactly this — reduced cognitive load as a selling point rather than a warning sign. The officer’s signature then attests, under penalty of perjury, to a document assembled by a machine.
    • Audit trails designed to limit exposure. Some tools log only the initial draft request and the officer’s sign-off, making it impossible to tell what the officer wrote and what the model wrote. CDT observed that some vendors advertise this limitation as a feature to “reduce FOIA exposure” — an accountability mechanism marketed as an accountability problem.

    The surveillance dimension is newer and less regulated. Forbes obtained AI-generated reports from the Richmond, California police department covering school bullying calls, natural deaths, and graphic sexual assault descriptions — encounters where no crime occurred, involving minors and innocent parties. One report documented a teenage student who typed a SpongeBob SquarePants quote on his laptop, prompted a police response after a classmate misread it as a threat, and was swiftly cleared — the report noted the incident would be “retained for future pattern tracking.” Calli Schroeder, senior counsel at the Electronic Privacy Information Center, told Forbes: “This is making what was supposed to be an accountability mechanism another tool of public surveillance.”[1]


    Two States, Two Philosophies

    Legislatures began responding before the empirical record was even complete. Utah enacted the nation’s first law regulating AI in police report writing: SB 180, sponsored by Senator Stephanie Pitcher — a former prosecutor — signed in March 2025. It requires that any report drafted wholly or partly by generative AI carry a disclosure and that the officer certify the accuracy of the final document, while leaving each agency to set its own implementation policy.[5]

    California followed with a more demanding model. SB 524, authored by Senator Jesse Arreguín and signed October 10, 2025, adds enforcement-relevant architecture that Utah’s law leaves to agency discretion:[6]

    • Disclosure on every page. Each report must prominently identify every AI program used and state: “This report was written either fully or in part using artificial intelligence.” The placement rule exists because disclosures buried at the end of a report function differently than ones a prosecutor sees on page one.
    • A signed accuracy certification. The preparing officer must sign, physically or electronically, verifying they reviewed the contents and that “the facts contained in the official report are true and correct” — converting the human-in-the-loop promise into a legal attestation.
    • Retention of the first draft. The machine’s original output must be preserved for as long as the final report, and — except for the official report itself — an AI draft “shall not constitute an officer’s statement.” This gives judges and defense attorneys a before-and-after record of what the model wrote versus what the officer kept.
    • An audit trail with named inputs. Agencies must log who used the AI and what body-camera video or audio the report drew from, so the draft can be checked against its source.
    • Vendor data restrictions. Contractors may not share, sell, or repurpose the information agencies feed into the system, closing a commercial surveillance loophole.

    Arreguín framed the bill in constitutional terms: “We’re not going to gamble with personal liberty. AI hallucinations happen at significant rates, and what goes in a police report can influence whether or not the state takes away someone’s freedom.” Kate Chatfield of the California Public Defenders Association made the due process case: “Too often, police narratives become the default truth, even when they contain errors or biases.”[7]

    The two laws represent a live philosophical split that other states are now watching. The R Street Institute’s Logan Seacrest argues Utah’s light-touch approach “provides structure without strangling innovation,” warning that inflexible mandates impose compliance costs on short-staffed departments and could chill experimentation while the technology matures. He credits both states with keeping a human in the loop, while cautioning that automation bias can still lull officers into deference — pointing to the Tennessee woman who spent five months in a North Dakota jail on an erroneous facial-recognition match. Police unions in California raised the same administrative-burden objections before the bill was narrowed by amendment.[8]


    The Institutional Countermeasures

    While legislatures debate, the actors who actually receive police reports have begun imposing their own rules — and because prosecutors hold the charging power, their policies function like law.

    • King County, Washington (2024). The Prosecuting Attorney’s Office in Seattle — which handles all prosecutions in the county — instructed police it would not accept narratives produced with AI assistance. Chief Deputy Prosecutor Daniel J. Clark’s memo explained the reasoning: “While an officer is required to edit the narrative and assert under penalty of perjury that it is accurate, some of the [genAI] errors are so small that they will be missed in review.” Interviewed by Forbes in 2026, Clark said he had seen the tools get witness and officer names wrong and place officers at scenes when they were only on the radio: “It’s a type of error that can really challenge an officer’s credibility when they’re on the stand, and then that puts the officer at risk of being either called lazy, sloppy or dishonest. None of those are good options.”[9]
    • Connecticut (April 2026). Chief State’s Attorney Patrick J. Griffin imposed a statewide moratorium on the use of AI programs “to draft, author and/or narrate” police reports — the first statewide prosecutorial freeze, covering all state’s attorneys. The move treats AI-drafted reports as an unresolved evidentiary risk rather than a disclosure problem: if the document cannot be vouched for, it does not matter that its use is labeled.[10]
    • The Daubert frontier. Guidance from the Prosecutors’ Center of Excellence suggests AI-generated reports should clear the Daubert reliability standard courts apply to scientific evidence — a bar that vendors guarding proprietary technical details may struggle to document. And within the police-tech industry itself, dissent is emerging: Truleo CEO Anthony Tassone, whose own company markets a report tool that assists rather than drafts, told Forbes that AI-crafted drafts are “worse than from starting [with] a blank page, because now you’ve given police this corrupted report.”[1]

    The logic uniting these responses is not hostility to AI; King County’s memo states plainly, “We do not fear advances in technology.” It is that the report’s function in the justice system is to be a verified human attestation, and a document assembled by a probabilistic model does not become one just because an officer clicks approve.[9]


    The Federal Evidence Question

    The federal rules of evidence are being rebuilt for exactly this problem. Proposed Federal Rule of Evidence 707, drafted by the Advisory Committee on Evidence Rules, states: “When machine-generated evidence is offered without an expert witness and would be subject to Rule 702 if testified to by a witness, the court may admit the evidence only if it satisfies the requirements of Rule 702(a)-(d).” In plain terms, machine output offered as proof would face the same Daubert-style gatekeeping as expert testimony, with a carve-out for “simple scientific instruments” like radar guns and breathalyzers.[11]

    The proposal drew heavy criticism during its public comment period, which closed February 16, 2026. The American Association for Justice urged the committee to pause and redraft, arguing the rule as drafted would sweep in routinely admitted evidence — geolocation data, surveillance footage, electronic health records — and recommending it be narrowed to machine-learning systems with a clear definition of “machine-generated.”[11]

    The committee has not been quick to act. At its May 7, 2026 meeting, it declined to recommend the rule, revised the draft — including deleting the simple-scientific-instruments exception in favor of a narrower judicial-notice carve-out, a change the Department of Justice opposed — and deferred further work to a fall 2026 mini-conference. Under the Rules Enabling Act pipeline, the earliest a machine-generated-evidence rule could take effect is December 1, 2027. Until then, the gatekeeping burden falls on the state disclosure laws and the prosecutors’ acceptance policies — the patchwork that is already operating.[12]


    What Both Sides Get Right

    The case for the tools is not fictional, and dismissing it misses why departments keep buying. Avon, Colorado Police Chief Greg Daly, who uses Draft One alongside two other products, described a genuine operational benefit: “To come into the station and have draft reports or the skeleton of a report to work on is extremely helpful and has a very beneficial mental health aspect.” Reports may also get written sooner after an encounter, when details are fresh, and some encounters may produce a written record that otherwise would not have existed at all.[1]

    But the certification architecture being built around these tools rests on an assumption the blinded research has now falsified: that a reviewing human can tell machine text from officer text and will catch what the machine got wrong. When 92 supervisors with two decades of experience detected AI authorship at the rate of a coin flip, and approved AI drafts at the same rate as human ones, the “human in the loop” stops being a safeguard and becomes a formality — a signature converting unverified machine output into sworn testimony.[3]

    That is the core insight beneath the state laws and the moratoria: the choice is not between AI reports and human reports, but between disclosed, audited, draft-retained AI reports and invisible ones. California’s retention-and-audit model creates the before-and-after record that makes adversarial testing possible; Utah’s certification model at least puts the defense on notice; a prosecutor’s refusal shifts the burden back to the vendor to prove reliability. What none of them can do is make a machine’s draft true — which is why the jurisdictions measuring outcomes before they scale, rather than after, are the ones whose records will hold up in court.[8]


    References

    [1] Forbes — “Axon Says AI Police Reports Save Time. Public Records Show They Get Facts Wrong” by Thomas Brewster (July 22, 2026) [SECONDARY]

    [2] Adams, Ian T., Barter, Matt, McLean, Kyle, et al. — “No man’s hand: artificial intelligence does not improve police report writing speed.” Journal of Experimental Criminology (published October 2, 2024) [PRIMARY]

    [3] Adams, Ian T., Barter, Matt, McLean, Kyle, et al. — “A Good College Essay but a Bad Police Report: A Triple-Blind Expert Evaluation of AI-Assisted Police Reporting.” CrimRxiv (May 8, 2026) [PRIMARY]

    [4] Center for Democracy & Technology — “Automated Police Report Drafting” [ADVOCACY]

    [5] Utah State Legislature — “S.B. 180 Law Enforcement Usage of Artificial Intelligence” (signed March 25, 2025) [PRIMARY]

    [6] California Legislative Info — “SB-524 Law enforcement agencies: artificial intelligence” (Chapter 587, signed October 10, 2025) [PRIMARY]

    [7] California State Senate (Sen. Jesse Arreguín) — “Governor Signs Bill Requiring Law Enforcement to Disclose Use of AI in Police Reports” (October 13, 2025) [PRIMARY]

    [8] Governing — “The Risky Marriage of AI and Police Reports” by Logan Seacrest, R Street Institute (July 15, 2026) [SECONDARY]

    [9] Electronic Frontier Foundation — “Prosecutors in Washington State Warn Police: Don’t Use Gen AI to Write Reports” by Matthew Guariglia (October 17, 2024) [ADVOCACY]

    [10] Connecticut Division of Criminal Justice — “Hartford Courant Interviews Chief State’s Attorney Patrick J. Griffin About Moratorium on the Use of Artificial Intelligence Programs For Police Reports” (April 20, 2026) [PRIMARY]

    [11] American Association for Justice — “Formal Comment — Machine-Generated Evidence (FRE 707)” (February 16, 2026) [ADVOCACY]

    [12] Lucid Truth Technologies — “FRE 707 AI Evidence: What Defense Attorneys Must Know Before Adoption” (April 27, 2026; updated July 20, 2026) [SECONDARY]


    AI automated and generated content. Human reviewed.

  • The Courtroom Frontier: How a Deepfake Evidence Case, a National AI Framework, and Three State Laws Are Redrawing the Line Between AI and Criminal Justice

    The Courtroom Frontier: How a Deepfake Evidence Case, a National AI Framework, and Three State Laws Are Redrawing the Line Between AI and Criminal Justice

    From the first detected deepfake submitted as evidence in a U.S. courtroom to a landmark decision framework for criminal justice agencies and three state laws governing AI police reports, risk assessments, and lethal robots, the justice system is scrambling to govern AI before it governs justice.


    Exhibit 6C: The First Deepfake Caught in Court

    On September 9, 2025, Judge Victoria Kolakowski of California’s Alameda County Superior Court did something no American judge had done before: she threw out a case because a piece of evidence was an AI-generated deepfake. The video at issue, labeled Exhibit 6C, purported to show a real witness speaking in a housing dispute. But the witness’s voice was “disjointed and monotone,” her face “fuzzy and lacking emotion,” and every few seconds she would “twitch and repeat her expressions.” Kolakowski determined the video had been produced using generative artificial intelligence, dismissed the case, and recommended sanctions against the plaintiffs for what she characterized as submitting false evidence.[1]

    The case, Mendones v. Cushman & Wakefield, Inc., appears to be one of the first known instances in which a deepfake was deliberately submitted as purportedly authentic evidence in a U.S. court and detected as AI-generated. The plaintiffs sought reconsideration, arguing the judge suspected but failed to prove the evidence was AI-generated. Kolakowski denied the request on November 6.[1]

    What makes the case especially significant is not just the deception itself but how it was uncovered. The court found that the metadata of one of the deepfaked videos showed it was captured on an iPhone 6 — impossible, given that the plaintiffs’ argument required capabilities only available on an iPhone 15 or newer. The technology to fabricate courtroom evidence has outpaced the legal system’s ability to detect it, and the metadata gap was almost an accidental catch.[1]

    Judge Stoney Hiljus of Minnesota’s 10th Judicial District, who chairs the Minnesota Judicial Branch’s AI Response Committee, put the stakes plainly: “I think there are a lot of judges in fear that they’re going to make a decision based on something that’s not real, something AI-generated, and it’s going to have real impacts on someone’s life.”[1]


    The Liar’s Dividend and the Deepfake Defense

    The Mendones case represents one side of the AI evidence problem — fabricated content admitted as real. But legal scholars are equally concerned about the reverse: real evidence dismissed as fake. This phenomenon, known as the “Liar’s Dividend,” occurs when parties invoke the mere possibility of AI manipulation to cast doubt on authentic evidence. In a 2023 lawsuit against Tesla over a self-driving crash, the company’s defense counsel unsuccessfully attempted to dismiss a video by claiming it was a deepfake.[1]

    A November 2025 report from the University of Colorado Boulder’s Visual Evidence Lab, compiled by 20 experts from around the country, frames both threats as part of a single crisis. More than 80% of court cases now hinge to some degree on video evidence, yet courts at both the state and federal levels lack clear guidelines for evaluating AI-enhanced or AI-generated footage. The report calls for specialized training for judges and jurors to critically evaluate visual evidence, national standards governing what kinds of AI are permissible in court, and a new system for storing and retrieving evidentiary videos.[2]

    Sandra Ristovska, the report’s senior author and director of CU’s Visual Evidence Lab, warned that the “deepfake defense” could be even more corrosive than fabricated evidence itself: “If this continues, jurors will accord little or no weight to authentic footage that they really should be paying attention to.”[2]

    The report also highlights an access-to-justice dimension. Lawyers increasingly submit authentic videos enhanced by AI to improve sound or visual clarity, but not everyone can afford those technologies. Judges vary in whether they allow AI-enhanced evidence at all. “There is a real concern that AI enhancement may exacerbate already existing inequalities in access to justice,” Ristovska said.[2]

    The Provenance Problem

    The CU Boulder report is part of a growing institutional response. The National Center for State Courts, in partnership with the Thomson Reuters Institute, has formed a consortium that labels deepfakes as “unacknowledged AI evidence” to distinguish them from “acknowledged AI evidence” like AI-generated accident reconstruction videos that all parties recognize as synthetic. The consortium published a cheat sheet for judges advising them to ask those presenting potentially AI-generated evidence to explain its origin, reveal who had access, share whether it was altered, and look for corroborating evidence.[1]

    Judge Erica Yew of California’s Santa Clara County Superior Court noted the absence of any centralized tracking system. “I am not aware of any repository where courts can report or memorialize their encounters with deep-faked evidence,” she said. “I think AI-generated fake or modified evidence is happening much more frequently than is reported publicly.”[1]


    A National Framework: The CCJ Task Force on AI

    While courts grapple with fabricated evidence, the broader question of how criminal justice agencies should evaluate, procure, and deploy AI tools is being addressed at the institutional level. In March 2026, the Council on Criminal Justice (CCJ) Task Force on Artificial Intelligence released Assessing AI for Criminal Justice: A User Decision Framework, a detailed guide designed to help law enforcement, courts, and corrections agencies evaluate AI tools before and after adoption.[3]

    The Task Force, launched in June 2025, is a national, nonpartisan initiative whose 15 members include AI technology developers, police executives, civil rights advocates, community leaders, and formerly incarcerated people. Its work is supported by researchers at RAND. Former Texas Supreme Court Chief Justice Nathan Hecht, who chairs the Task Force, framed the urgency: “Thousands of criminal justice agencies are looking at how AI can improve their performance, but most are flying blind. There’s no objective guidance about how to maximize the potential benefits and minimize the possible harms.”[3]

    The framework’s core logic is a structured classification process. It begins by assessing whether an agency has sufficient staff capacity and technical expertise to deploy AI at all — a prerequisite often overlooked in the rush to adopt new tools. Technologies are then categorized based on their level of risk to due process, civil rights, and civil liberties, as well as their potential to enhance efficiency and crime prevention. Lower-risk tools may proceed through standard deployment; higher-risk systems require enhanced safeguards, further evaluation, or abandonment entirely. Ten appendices provide practical tools including readiness assessments, protocols for prohibited systems, system complexity evaluations, sector-specific guidance, procurement checklists, and ongoing monitoring templates.[3]

    The framework draws an important distinction between purpose-built AI acquired through formal procurement — like a police department’s report-writing software — and general-purpose AI tools like chatbots that staff may use informally for research or document drafting. The Task Force recommends that agencies develop specific policies governing staff use of general-purpose AI for case-related work, recognizing that the governance challenge extends well beyond purchased systems.[3]

    Yet the Task Force was candid about the limits of self-regulation. “This framework provides a structured pathway for responsible AI adoption in criminal justice, but frameworks alone are not sufficient to guarantee good outcomes,” the group said. “Important questions remain about what infrastructure is needed to make the full recommendations embedded in this framework accessible and considerate of the ways that AI uses may evolve.”[3]


    Three State Laws, Three Different Risks

    While the CCJ framework provides voluntary guidance, state legislatures have begun enacting binding law. Three enacted bills from 2025 illustrate the range of AI risks in criminal justice — from AI-generated police reports that can distort the evidentiary record, to algorithmic risk assessments that can determine whether someone goes free or stays locked up, to autonomous robots that could one day use lethal force.

    Utah: Disclosing AI-Generated Police Reports

    Utah’s S.B. 180, enacted in 2025, requires every law enforcement agency in the state to maintain a policy governing its use of generative AI. The law mandates that any police report or law enforcement record created wholly or partially using generative AI must include a disclaimer disclosing that fact. The report’s author must certify that they have read and reviewed the document for accuracy.[4]

    The law is a direct response to the rapid spread of products like Axon’s Draft One, which uses body-worn camera audio to generate police narrative reports automatically. The Electronic Frontier Foundation (EFF), an advocacy organization, supported the bill but warned that a disclaimer alone is insufficient. EFF raised concerns about AI’s ability to accurately process dialects, vernacular, and slang — noting that in a tense setting like a traffic stop, “AI mistaking a metaphorical statement for a literal claim could fundamentally change the content of a police report.” The organization also warned that AI-generated reports could give officers who deliberately exaggerate or lie a “veneer of plausible deniability,” since they could blame the AI for mischaracterizing body camera footage.[5]

    The problem is not hypothetical. Prosecutors in King County, Washington (which includes Seattle) have already instructed officers not to use generative AI to write reports, out of fear that the technology would compromise criminal cases.[5]

    Virginia: Human Oversight for AI Risk Assessments

    Virginia’s HB 1642, enacted in 2025, takes aim at a different category of risk: algorithmic tools that inform decisions about pretrial release, sentencing, probation, and parole. The law provides that recommendations or predictions from any AI-based tool used for criminal justice purposes “shall be reviewed and approved by a qualified human decision maker prior to using any such recommendations or predictions in making or informing a final decision.”[6]

    The law’s core principle is absolute: “No decision related to pretrial release, sentencing, probation, or parole may rely solely on the recommendation of an artificial intelligence-based tool without documented human oversight.” Documentation must include the AI-generated recommendation and the human decision maker’s rationale for accepting, modifying, or rejecting it.[6]

    This requirement addresses one of the most consequential applications of AI in criminal justice. Risk assessment algorithms have been used for over a decade to inform decisions about whether defendants are released before trial or granted parole. The Wisconsin Supreme Court’s 2016 decision in State v. Loomis upheld the use of the COMPAS risk assessment tool even though its proprietary algorithm could not be independently examined — a ruling that highlighted the tension between algorithmic opacity and due process. Virginia’s law does not ban these tools, but it ensures that a named human being must take documented responsibility for every decision they inform. The approach preserves the efficiency benefits of algorithmic analysis while creating an accountability anchor that the Loomis court struggled to locate.[6]

    North Dakota: Banning Autonomous Lethal Force

    North Dakota’s HB 1613, enacted in April 2025, addresses perhaps the most extreme AI risk in law enforcement: autonomous lethal force. The law provides that a law enforcement agency “may not authorize the use of a robot to deploy a lethal weapon, unless the weapon is controlled remotely, cannot be activated autonomously, and is used to neutralize an inanimate object at risk of exploding or causing an immediate threat to public safety or property.”[7]

    The law draws a clear line: robots may carry weapons, but a human must control the trigger, and the system must be incapable of autonomous activation. The narrow exception for explosive devices reflects a practical recognition that some threats require immediate action, but even there, the human-remote-control requirement holds. This is one of the first state laws to explicitly regulate the intersection of AI, robotics, and lethal force in policing — a domain that has been largely governed by internal agency policy rather than statute.[7]


    The State Sovereignty Argument

    The state-level legislative activity comes against a backdrop of federal pressure. In December 2025, the Trump administration issued an executive order aimed at curbing state AI regulation, directing the Department of Justice to sue states with laws deemed “onerous,” threatening to strip federal funding, and directing federal agencies to assert preemption over state AI governance. The order’s preemption provisions, however, specifically exempt laws regulating “state government procurement and use of AI” — meaning laws like Utah’s, Virginia’s, and North Dakota’s, which regulate their own agencies’ use of AI, fall outside its scope.[8]

    A April 2026 Brookings Institution analysis by Chiraag Bain, Alex Chohlas-Wood, and Katie Kinsey — two of whom serve on the CCJ AI Task Force — argued that states should not be deterred regardless. The analysis contends the executive order “has no preemptive power whatsoever” because “an executive order is not a law. The president cannot unilaterally override state statutes.” The authors note that the contemplated DOJ lawsuits face “serious legal obstacles” under the dormant commerce clause, the funding threats conflict with the Constitution’s spending clause requirements, and the FTC and FCC lack statutory authority to broadly preempt state AI law.[8]

    The Brookings authors point to a bipartisan consensus among governors that states should retain regulatory authority. California Governor Gavin Newsom, Utah Governor Spencer Cox, and Florida Governor Ron DeSantis — spanning the political spectrum — have all asserted their states’ right to regulate AI. Cox, a Republican, argued that “the Supreme Court is going to back me up on that.” DeSantis declared “we have a right” to pass AI regulation just weeks after Trump signed the order.[8]

    The underlying logic is straightforward: criminal justice is a domain of inherent state authority, and AI tools deployed in that domain — from police reports to risk assessments to robots — affect liberty and life in ways that demand democratic accountability. As the Brookings analysis puts it, state-level regulation is needed “to ensure the safety and fairness of this technology,” and if done right, can “accelerate the responsible and effective adoption of AI tools shown to improve public safety while protecting civil rights and liberties.”[8]


    The Governance Gap

    The developments of 2025 and 2026 reveal a criminal justice system at an inflection point. The Mendones case demonstrated that AI-generated evidence is no longer a theoretical threat — it has already entered a courtroom and deceived (briefly) the fact-finding process. The CU Boulder report showed that the legal system lacks the infrastructure, training, and standards to distinguish authentic from fabricated visual evidence at scale. The CCJ framework offered a voluntary pathway for responsible AI adoption but acknowledged that “frameworks alone are not sufficient.”[3]

    The state laws — Utah’s disclosure mandate, Virginia’s human oversight requirement, North Dakota’s lethal force prohibition — each address a specific, concrete risk. But they also reveal the patchwork nature of the response. A suspect in Virginia has the right to human-reviewed risk assessments; a suspect in a state without such a law may not. A police report in Utah must disclose AI involvement; in most other states, it need not. A robot in North Dakota cannot autonomously deploy lethal force; in states without similar statutes, the question is left to agency policy.

    Several pending bills suggest the patchwork may be narrowing. New York’s A.7172 would require the state’s Division of Criminal Justice Services to formulate a protocol for regulating AI and facial recognition technology in criminal investigations and restrict the use of AI-generated outputs in court. New York’s A.1338 would set rules for the admissibility of evidence created or processed by AI. California’s S.B. 524 would require law enforcement agencies to maintain policies for AI-generated reports, similar to Utah’s approach. These bills, if enacted, would begin to establish a common floor for AI governance in criminal justice across multiple jurisdictions.[9]

    The question is whether that floor will be built in time. As Judge Scott Schlegel of Louisiana’s Fifth Circuit Court of Appeal — a leading advocate for judicial AI adoption — put it: “My wife and I have been together for over 30 years, and she has my voice everywhere. She could easily clone my voice on free or inexpensive software to create a threatening message that sounds like it’s from me and walk into any courthouse around the country with that recording.” His conclusion was stark: “The judge will sign that restraining order. They will sign every single time.”[1]


    References

    [1] NBC News — “AI-generated evidence showing up in court alarms judges” by Jared Perlo (November 18, 2025) [SECONDARY]

    [2] University of Colorado Boulder — “Deepfakes and AI in the courtroom: Report calls for legal reforms to address a troubling trend” by Lisa Marshall (November 17, 2025) [SECONDARY]

    [3] Council on Criminal Justice — “National Task Force Releases New Framework to Help Criminal Justice Agencies Assess AI Tools” (March 31, 2026) [PRIMARY]

    [4] Utah State Legislature — “S.B. 180 Law Enforcement Usage of AI” (2025, Enacted) [PRIMARY]

    [5] Electronic Frontier Foundation — “Utah Bill Aims to Make Officers Disclose AI-Written Police Reports” by Matthew Guariglia (February 21, 2025) [ADVOCACY]

    [6] Virginia Legislative Information System — “HB 1642: Criminal Justice Risk Assessments; Human Oversight” (2025, Enacted) [PRIMARY]

    [7] North Dakota Legislative Assembly — “HB 1613: Law Enforcement Use of a Robot” (2025, Enacted) [PRIMARY]

    [8] Brookings Institution — “States can—and should—regulate AI in criminal justice” by Chiraag Bain, Alex Chohlas-Wood, and Katie Kinsey (April 16, 2026) [SECONDARY]

    [9] National Conference of State Legislatures — “Summary of Artificial Intelligence 2025 Legislation” (2025) [PRIMARY]


    AI automated and generated content. Human reviewed.

  • The Governance Gap: How Two Police Report Laws, Fifteen Facial Recognition Statutes, and a Federal Preemption Threat Are Reshaping AI in Criminal Justice

    The Governance Gap: How Two Police Report Laws, Fifteen Facial Recognition Statutes, and a Federal Preemption Threat Are Reshaping AI in Criminal Justice

    State legislatures are building the first guardrails around police AI — from California’s audit-trail mandate to Virginia’s new warrant requirement — while scholars warn that ad hoc rules are no substitute for dedicated governance institutions.


    The Speed Problem

    Artificial intelligence is no longer a distant or speculative technology in the criminal justice system. It is becoming part of its everyday machinery. Police departments use algorithmic tools to analyze digital evidence, look for patterns in crime data, and draft police reports. Prosecutors rely on AI software to manage discovery and support charging decisions. Courts encounter algorithmic risk assessments and large language models that promise to summarize records, draft documents, and assist with legal analysis[1]. Across the system, AI is being woven into decisions that affect liberty itself.

    What is changing is the speed, scope, and complexity of those tools. As police departments accumulate growing volumes of digital evidence — from body camera footage and surveillance video to jail calls, social media records, and case files — AI is increasingly being used to sort, search, and analyze that information. Cris Moore, a computer scientist and professor at the Santa Fe Institute, told Stateline that “the speed at which technologically created evidence has been adopted, and the aggression with which it’s being pushed makes it hard for the legal community to keep up”[2].

    The companies offering AI-powered tools for law enforcement include Axon, Motorola Solutions, TRULEO, Flock Safety, Clearview AI, and Mark43, whose ReportAI and BriefAI products help officers draft reports from dispatch records and body camera footage. Mark43’s senior vice president of product, Wendy Gilbert, framed the value proposition: “Our core mission is to help responders spend less time on administrative work, so that they can spend more time serving in their communities”[2]. Yet the same efficiency gains that make these tools attractive to under-resourced agencies also make them harder to vet. A May 2026 RAND report produced for the Council on Criminal Justice’s Task Force on AI found that AI applications are “often discussed and governed as if they were a single category of technology,” despite wide variation in their technical sophistication, intended use, and influence on human judgment[3].


    Two States, Two Models: California and Utah Write the First Rules for AI Police Reports

    As of July 2026, only two states have enacted laws specifically governing AI-assisted police report writing. Both require disclosure and officer certification, but they differ in how deeply they reach into the audit trail — and that difference reveals two philosophies about what oversight is for.

    California SB 524 — The Comprehensive Model

    Enacted and effective January 1, 2026, California’s SB 524 is the most comprehensive law of its kind in the country. When a report is generated wholly or partly by AI, the agency’s policy must ensure[4]:

    • Disclosure of the AI program used — the specific system that generated or assisted the report must be identified, so that downstream actors (prosecutors, defense counsel, judges) know what they are dealing with.
    • Officer certification of accuracy — the officer must affirm that they reviewed the report for accuracy, establishing that the final report is the officer’s statement, not the machine’s.
    • Retention of the original AI-generated draft — agencies must keep the first AI-generated draft for as long as the final report is kept, preserving a paper trail that can reveal what the AI said before a human edited it.
    • An audit trail tied to source footage or audio — the system must link outputs back to the body camera or dispatch audio they were generated from, so that claims in the report can be checked against raw evidence.
    • A prohibition on vendors reselling agency data — police data collected through AI report-writing cannot be monetized or repurposed by the vendor.

    The core logic here is evidentiary integrity. Police reports play a critical role in investigations and court proceedings, and errors introduced by AI systems — inaccuracies, omissions, or misinterpretations of context — could influence how evidence is understood by investigators, prosecutors, and judges if they go undetected. By mandating draft retention and source-linking, California ensures that the gap between what the machine produced and what the officer signed off on is always reconstructable. That is the difference between a report that says “an AI helped” and one that can be forensically examined.

    Utah SB 180 — The First-Mover Model

    Utah was the first state to act, with SB 180 effective May 7, 2025. It requires a written generative-AI policy, a disclaimer on any AI-assisted report, and officer certification of accuracy[4]. Utah does not require retaining the original AI draft — a meaningful gap compared to California, because without the draft there is no way to see what the AI initially produced or whether the officer’s edits changed substantive content or merely corrected formatting. The Utah model trusts the certification; the California model trusts but verifies.

    The Bills That Did Not Make It

    Several other states tried and failed. New York’s S10425 / A9253-A was the most ambitious bill — its “Covered AI” definition expressly included generative report-writing AI, with a public inventory, disclosure, first-draft retention, and both Attorney General and private enforcement — but it died when the Legislature adjourned sine die on June 5, 2026. Reintroduction is expected in 2027[4]. Minnesota’s SF 4575 / HF 4536, which would have prohibited generative AI in official records outright, remains pending. Virginia’s HB 1294 was carried over to 2027. Nevada’s SB 199, a proposed ban, failed in 2025[4].

    In the absence of statutes, some prosecutors have stepped in. Connecticut’s Chief State’s Attorney imposed a statewide moratorium on using AI to draft or narrate criminal reports in April 2026. King County, Washington’s Prosecuting Attorney’s Office will not accept AI-assisted report narratives, per a September 2024 memo[4]. These non-statutory restrictions show that the gap between legislative action and frontline practice is being filled — unevenly — by prosecutors who answer to no legislature.


    The Facial Recognition Frontier: 15 States and a Virginia Tightening

    Facial recognition is the most regulated AI application in law enforcement, but the regulatory landscape is fractured across at least four distinct approaches. Roughly 15 states have enacted limits as of July 2026, and the strategies cluster into recognizable patterns[4]:

    • Warrant or court-order requirements (Washington, Montana, Massachusetts, Maryland, Utah, and as of July 1, 2026, Virginia) — facial recognition data can only be accessed through a judicial warrant, treating it like a search under the Fourth Amendment rather than an investigative convenience.
    • “Not the sole basis for arrest” safeguards (Alabama, Maryland, Montana, Washington) — a face match alone cannot justify an arrest; it must be corroborated by independent evidence. This directly addresses documented wrongful arrests, primarily of Black individuals, caused by erroneous facial recognition identifications[5].
    • Near-total moratoria (Maine, with the broadest government ban and narrow carve-outs; Vermont, with only a child-exploitation exception) — these states have decided the risks of police facial recognition outweigh the benefits entirely.
    • Body-camera-only prohibitions (Minnesota, Oregon, New Hampshire) — barring facial-recognition analysis of body-cam footage specifically, preventing the tool that officers wear every shift from becoming a continuous biometric surveillance system.

    Virginia’s July 1 Tightening

    Virginia’s evolution is instructive because it shows how a state can move from permissive to restrictive within a single legislative cycle. The state’s prior law, still effective until July 1, 2026, allowed local law enforcement to use facial recognition for a long list of “authorized uses” — from identifying crime victims to vetting undercover officers — and required NIST evaluation for accuracy and demographic performance[6]. It was a permissive-but-regulated model.

    The new version, effective July 1, 2026, takes a fundamentally different approach. It provides that “no local law-enforcement agency shall purchase or deploy facial recognition technology unless such purchase or deployment of facial recognition technology is expressly authorized by statute”[6]. A statute that does not refer to facial recognition technology cannot be construed to provide that authorization. The law further requires that any facial recognition technology be “maintained under the exclusive control of such local law-enforcement agency” and that any data be “kept confidential, not be disseminated or resold, and be accessible only by a search warrant”[6].

    The logic is a default-deny posture: unless a locality’s governing body has specifically voted to authorize facial recognition with warrant-level data protections, the technology cannot be used. This flips the burden from “allowed unless prohibited” to “prohibited unless expressly authorized” — a structural choice that makes it much harder for agencies to adopt the technology quietly through procurement.

    Where Courts Step In

    In states without statutes, courts have begun filling the gap. In New Jersey, the case State v. Arteaga (2023) requires defendants be notified when facial recognition was used in their case. In Michigan, a settlement barred the Detroit Police Department from arresting on a face match alone[4]. These judicial interventions matter because they create binding rules in individual jurisdictions without legislative action — but they also produce a patchwork where your rights depend on your zip code.


    The Preemption Shadow: Can the Federal Government Stop States?

    While states build these guardrails, a federal executive order threatens to chill the entire enterprise. In December 2025, the White House released a presidential action titled “Eliminating State Law Obstruction of National Artificial Intelligence Policy,” designed to intimidate state AI laws through legally dubious means[5].

    The order deploys the Department of Justice to sue states, threatens to strip federal funding from states with laws deemed “onerous,” directs the Federal Trade Commission to declare state algorithmic discrimination laws preempted, and instructs the Federal Communications Commission to create reporting standards that would nullify state transparency requirements[5]. Crucially, the order itself exempts from its scope any laws regulating “state government procurement and use of AI” — meaning states retain clear authority to regulate their own law enforcement and criminal justice agencies’ use of AI[5].

    Brookings scholars Chiraag Bains, Alex Chohlas-Wood, and Katie Kinsey argue in an April 2026 analysis that the administration’s main legal theory — that state AI laws violate the dormant commerce clause — is weak, because most state AI laws apply equally to in-state and out-of-state developers. They also note that the president “cannot by executive fiat attach new policy conditions to BEAD” funding and that nothing in the FCC’s statutory authority authorizes it to regulate AI generally, let alone preempt state AI laws[5]. Their bottom line: “States should not be deterred”[5].

    Utah’s Republican Governor Spencer Cox has separately argued that his state should be allowed to regulate AI to avoid harms, noting that “the Supreme Court is going to back me up on that”[5]. The preemption fight is not partisan in the traditional sense — it pits a federal deregulatory agenda against state authority that both red and blue states have been exercising.


    Beyond Statutes: The Case for a Dedicated Governance Entity

    Even where states have acted, the rules are narrow and reactive. California regulates AI in police reports. Virginia regulates facial recognition. But no state has built a governance institution capable of evaluating the full range of AI tools flowing into criminal justice — from predictive policing models to generative AI in prosecutorial charging decisions to algorithmic risk assessments at sentencing.

    A March 2026 Stanford Law School policy paper, produced in partnership with the Council on Criminal Justice’s Task Force on AI, argues that this is the central problem. The paper’s core finding is that “artificial intelligence capabilities and products are being developed and deployed in criminal justice without sufficient understanding of how they work, attention to their failure risks, or analysis of how they implicate constitutional and other values”[1]. Judges, prosecutors, police officers, probation officers, and policymakers are adopting complex technologies without the technical background needed to assess their design, limitations, or possible failures. Vendors market these systems directly to practitioners, most of whom lack both the time and expertise to conduct meaningful independent evaluation[1].

    The Stanford team’s recommendation is structural: we need specialized AI governance entities for the criminal justice system. Such an entity — housed within the federal government, a consortium of states, or a public-interest institution with formal authority and stable funding — could serve as a hub of technical, legal, and ethical expertise. It could independently evaluate high-impact AI tools, publish accessible assessments, maintain up-to-date guidance on acceptable uses and known risks, develop standardized benchmarks and testing protocols, coordinate incident reporting and auditing, and provide technical assistance to agencies that cannot afford their own AI review teams[1].

    The paper draws an analogy to sentencing commissions, which it identifies as the closest existing model. Sentencing commissions balance expertise, stability, transparency, and iterative governance. They pair pluralistic decision-making bodies with standing research staff, translating technical analysis into durable standards that shape practice over time, even without broad enforcement authority. The Stanford team concludes that “the most effective approach to AI governance in criminal justice may draw selectively from multiple models rather than replicating any single one wholesale”[1].

    The underlying principle is that good guidelines and standards are necessary but not sufficient. The CCJ Task Force is doing vital work in articulating governing principles, procurement standards, and case studies. The question is how to deploy and implement that guidance — how to build durable structures that can translate expert recommendations into concrete policies and support practitioners in applying them case by case, tool by tool[1]. Without that institutional capacity, standards on paper become like the Utah draft-retention gap: well-intentioned rules that cannot be enforced or even checked.


    The Agentic Policing Horizon

    The most forward-looking concern raised by experts is the possible advent of what Andrew Guthrie Ferguson, a law professor at George Washington University and author of “Your Data Will Be Used Against You: Policing in the Age of Self-Surveillance,” calls “agentic policing.” Future technologies could integrate body-camera footage, camera networks, and other data sources into a single system capable of generating investigative leads, identifying potential suspects, or suggesting connections between cases[2].

    Ferguson’s concern is not that AI will replace human judgment but that it will invert the investigative process. “All that data is going to be dumped into an AI model, and they’re going to query it to say who’s the most likely suspect,” he told Stateline. “The AI is going to be running the agentic analysis of it and come up with the answer, and then police and prosecutors have to kind of work backwards to see if it’s accurate”[2]. This flips the traditional investigative process — which starts with evidence and builds toward a conclusion — into one that starts with a conclusion and works backward to justify it. Ferguson warns: “We’ve never started with an answer and made people work backwards. There are very real constitutional, statutory and practical risks with this new model of agentic policing”[2].

    The RAND taxonomy for the CCJ Task Force reinforces this concern with a structural finding: risks for complex AI systems are “concentrated in high-stakes criminal justice functions” including pretrial, sentencing, and enforcement decisions, and AI applications relying on past criminal justice data “tend to systematically reproduce racial and socioeconomic disparities”[3]. The taxonomy recommends that agencies “prioritize safeguards over expansion in high-risk areas” and “keep humans in charge of decisions about liberty”[3].


    Conclusion: The Patchwork Problem

    The current moment is defined by a paradox. AI adoption in criminal justice is accelerating — Mark43 alone reports that dozens of agencies are using, testing, or evaluating its AI features, and Hawaii’s Maui County Council approved a $1.7 million expansion of AI-supported policing tools in June 2026[2] — but the governance infrastructure is being built state by state, statute by statute, with no national framework and no consistent standard for what “responsible AI in criminal justice” means in practice.

    California’s SB 524 sets the high-water mark for police report oversight. Virginia’s July 1 tightening shows that states can move from permissive to restrictive when the evidence warrants. The CCJ Task Force and Stanford have built the intellectual architecture for a governance entity that could evaluate tools, publish assessments, and provide technical assistance. But none of these pieces, by itself, closes the gap. As the Stanford paper puts it: “The tools are here. The stakes are clear. What is missing is the governance capacity to match them. It is time to build it”[1].

    Until that capacity exists, the rules that govern AI in criminal justice will continue to be written in the space between a vendor’s marketing pitch and a legislature’s next session — and the people whose liberty is at stake will have to hope their state was one of the ones that showed up.


    References

    [1] Stanford Law School — “AI in Criminal Justice: Why Governance Matters and How to Make It Work” by Jonathan Wroblewski (March 27, 2026) — [SECONDARY]

    [2] Stateline — “Police use of artificial intelligence grows as rules lag behind” by Amanda Watford (June 26, 2026) — [SECONDARY]

    [3] Council on Criminal Justice — “An AI Taxonomy for Criminal Justice” (RAND, May 2026) — [SECONDARY]

    [4] Barricade AI — “AI Police Report Laws by State: 2026 Tracker” (July 29, 2026) — [SECONDARY] (advocacy organization — Barricade AI builds AI products for law enforcement and discloses a commercial interest)

    [5] Brookings Institution — “States can—and should—regulate AI in criminal justice” by Chiraag Bains, Alex Chohlas-Wood, and Katie Kinsey (April 16, 2026) — [SECONDARY]

    [6] Virginia Law — “Code of Virginia § 15.2-1723.2. Facial recognition technology; approval” (Effective July 1, 2026) — [PRIMARY]


    AI automated and generated content. Human reviewed.


    Listen to the Bears Discuss This

    Episode 15 of AI Policy for the Bears

  • The 315-Use-Case Reckoning: How Federal AI Expansion, a Preemption Executive Order, and a Private Surveillance Network Are Reshaping Criminal Justice Policy

    The 315-Use-Case Reckoning: How Federal AI Expansion, a Preemption Executive Order, and a Private Surveillance Network Are Reshaping Criminal Justice Policy

    Federal AI deployment in criminal justice jumped 31% in a single year, while a December 2025 executive order tries to block state guardrails and a private nonprofit quietly builds America’s first live facial recognition network.


    The Federal AI Expansion at Scale

    The Department of Justice’s 2025 AI Use Case Inventory logged 315 entries, a 30.7% increase from the 2024 inventory and a staggering leap from just four reported use cases two years prior.[1] Of those 315, 114 were classified as “high-impact,” a designation covering AI systems whose decisions or actions carry a “legal, material, binding or significant effect on rights and safety.”[2] The inventory spans every corner of the agency, from the Federal Bureau of Prisons using predictive AI to classify inmate misconduct risk, to the ATF analyzing airline travel data for “atypical” passenger routes, to the U.S. Marshals Service considering Amazon Rekognition for facial recognition.[2]

    What makes this expansion notable is not just its volume but the range of consequential decisions it touches. The Federal Bureau of Prisons’ “BRAVO Classification” system uses statistical techniques to predict the potential for misconduct among newly admitted inmates and assign security levels accordingly.[2] The DOJ’s Office of Justice also deployed an agentic AI tool called PATTERN, designed to predict recidivism risk for incarcerated adults and ostensibly reduce re-engagement with the justice system.[2] PATTERN is one of only four agentic AI use cases in the entire 2025 inventory, meaning it can take autonomous actions rather than merely generating outputs for human review, a category that did not exist in the prior year’s inventory at all.[2]

    Privacy and civil liberties advocates have raised sharp concerns about the trajectory. Jay Stanley, a senior policy analyst with the ACLU’s Speech, Privacy, and Technology Project, said the DOJ’s 2025 inventory provides a “snapshot” of how the federal government “is aggressively seeking to test and exploit a wide variety of AI algorithms and sifting through data on ordinary people.”[2] Jacob Hoffman-Andrews, a senior staff technologist at the Electronic Frontier Foundation, noted that tools predicting criminality or recidivism inevitably make inferences from correlated variables: “Even if you tell us specifically not to, AI is very good at learning correlated variables and figuring out how to be biased.”[2]


    The DOJ’s Own Roadmap for Guardrails

    Before the current administration’s policy shifts, the Department of Justice produced a 77-page final report on AI in criminal justice, submitted in December 2024 pursuant to the Biden-era Executive Order 14110.[3] The report examines four main applications: identification and surveillance, forensic analysis, predictive policing, and risk assessment, and offers detailed recommendations for each.[4] While the policy environment has evolved since its publication, the technical and ethical considerations it identifies remain the most comprehensive federal framework for responsible AI deployment in criminal justice.

    Identification and Surveillance

    The report covers facial recognition technology, automated license plate readers, and other biometric systems. Its key recommendations include mandatory training for all system users, clear policies on authorized use cases, quality control procedures, regular auditing and monitoring, and transparent public reporting.[4] The core logic is that these tools offer genuine efficiency gains but require comprehensive policies and oversight to prevent the documented harms of false matches and biased performance across demographic groups.

    Predictive Policing

    The report distinguishes between place-based tools, which forecast crime hotspots, and person-based tools, which attempt to identify individuals likely to commit or be victimized by crime. Its recommendations stress community engagement before adoption, careful data selection to avoid feedback loops, independent third-party auditing of models and source code, and periodic reevaluation of models based on policing patterns and crime rates.[3] The underlying principle is that predictive models may carry a “veneer of neutrality” reinforced by automation bias, making it essential that humans retain a central role in choosing data, evaluating outputs, and determining interventions.[3]

    Risk Assessment

    Risk assessment instruments, which use quantitative analysis to estimate recidivism likelihood or failure to appear for trial, are currently used in some form by every state.[3] The report’s recommendations address tool design, implementation, and validation, emphasizing that risk assessment output should inform, not replace, judicial discretion. The Council on Criminal Justice’s summary of the report notes that successful implementation requires agencies to establish robust organizational structures, ensure public oversight and transparency, and develop appropriately trained workforces, both before and after deployment.[4]


    The December 2025 Executive Order and the Preemption Push

    On December 11, 2025, President Trump signed Executive Order 14365, titled “Ensuring a National Policy Framework for Artificial Intelligence.”[5] The order frames state-level AI regulation as an obstacle to national competitiveness, arguing that “State-by-State regulation by definition creates a patchwork of 50 different regulatory regimes that makes compliance more challenging, particularly for start-ups.”[5] It specifically calls out a Colorado law banning “algorithmic discrimination,” suggesting it may “force AI models to produce false results in order to avoid a ‘differential treatment or impact’ on protected groups.”[5]

    The order establishes several mechanisms to challenge state AI laws:

    • AI Litigation Task Force: The Attorney General must establish a task force within 30 days to challenge state AI laws on grounds including unconstitutional interstate commerce regulation and federal preemption.[5]
    • Funding restrictions: The Secretary of Commerce must issue a policy notice specifying that states with “onerous” AI laws are ineligible for non-deployment funds under the BEAD broadband program, and other agencies must assess whether they can condition discretionary grants on states not enforcing such laws.[5]
    • FTC and FCC proceedings: The FTC must issue a policy statement on when state laws requiring alterations to AI model outputs are preempted by the Federal Trade Commission Act, and the FCC must initiate a proceeding to determine whether to adopt a federal reporting standard that preempts conflicting state laws.[5]

    However, the order contains a critical carve-out. Section 8(b) specifies that the legislative recommendation for a federal preemption framework should not propose preempting state AI laws relating to “child safety protections,” “AI compute and data center infrastructure,” and crucially, “State government procurement and use of AI.”[5] This means states retain explicit authority to regulate how their own law enforcement and criminal justice agencies procure and use AI systems.


    Why the Preemption Threat Rests on Legal Quicksand

    A April 2026 analysis by Chiraag Bains, Alex Chohlas-Wood, and Katie Kinsey, published by the Brookings Institution, argues that the executive order’s threats to state regulation are built on dubious legal ground.[6] The authors make several key points about why states should not be deterred:

    • No inherent preemptive power: An executive order is not a law. The president cannot unilaterally override state statutes. The order merely directs agencies to take actions that might eventually create pathways for preemption, and each pathway faces serious legal obstacles.[6]
    • Dormant commerce clause weakness: The administration’s main theory is that state AI laws violate the dormant commerce clause, which primarily bars economic protectionism. But most state AI laws apply equally to in-state and out-of-state developers, and AI companies can tailor software to different markets, making it difficult to show excessive burden.[6]
    • Spending clause limits: The president cannot attach new policy conditions to BEAD funding by executive fiat. The BEAD statute does not mention AI, and AI deregulation does not advance Congress’s goal of expanding high-speed internet access, so the conditions would likely fail the constitutional requirement that funding conditions relate to grant program purposes.[6]
    • FTC and FCC authority limits: The FTC cannot preempt state regulation by mere assertion, and nothing in the FCC’s statutory authority, which is limited to the communications industry, authorizes it to regulate AI generally or preempt state AI laws.[6]

    The authors conclude that the administration’s threats amount to intimidation designed to “discourage state policymakers from even trying to protect their constituents,” calling it “a direct assault on state sovereignty and violates basic principles of federalism.”[6] Both Republican and Democratic governors have signaled they will continue regulating AI regardless. California Governor Gavin Newsom cited the “failure of the federal government to enact comprehensive, sensible AI policy” when signing AI legislation. Utah Governor Spencer Cox argued his state should be allowed to regulate AI, noting that “the Supreme Court is going to back me up on that.” Florida Governor Ron DeSantis declared “we have a right” to pass AI regulation just weeks after Trump signed the order.[6]


    New Orleans: America’s First Live Facial Recognition Network

    While the federal-state regulatory battle plays out, a private nonprofit in New Orleans has quietly built what appears to be America’s first live facial recognition surveillance network. Project NOLA, founded in 2009 by former police officer Bryan Lagarde, operates a network of over 5,000 cameras mounted on private property by “volunteers” who pay annual connection fees.[7] About 200 of the network’s most advanced cameras have live facial recognition capability, constantly scanning faces in the French Quarter and comparing them against approximately 250 people on “hot lists.”[7]

    The system exists in a legal gray zone. New Orleans Police Superintendent Anne Kirkpatrick determined in April 2025 that cooperation with Project NOLA might violate a city ordinance that allowed facial recognition but not live facial recognition from a third-party entity, and she paused the department’s use of real-time alerts.[7] Yet Project NOLA continues feeding information to police, sending “unsolicited” tips when its cameras spot individuals the department is looking for.[7] Federal law does not specifically address how law enforcement may use live facial recognition, though the Supreme Court ruled in United States v. Jones that police must have probable cause and a warrant to use technology to follow someone continuously.[7]

    University of Washington law professor Ryan Calo described the arrangement as a potential “shell game” for legal responsibility, warning that “if the surveillance is done ‘by the community,’ by people that are not official actors, that it will circumvent those protections.”[7] The case illustrates how private actors can deploy AI surveillance technologies that police departments are legally or politically constrained from operating themselves, creating a workaround for constitutional protections.


    State and Local Guardrails Taking Shape

    Despite the federal preemption push, states and municipalities continue building guardrails for AI in policing. According to a Rutgers policy analysis, the pattern across jurisdictions reveals three consistent concerns driving restrictions: whether technologies actually improve outcomes, whether they perpetuate bias, and whether they maintain public legitimacy.[8]

    • Facial recognition bans and moratoria: San Francisco became the first major U.S. city to ban government use of facial recognition in 2019, followed by Boston and Portland. California imposed a temporary moratorium on facial recognition in police body cameras, while Washington and Colorado implemented requirements for accountability reporting, testing standards, and oversight before deployment.[8]
    • Predictive policing rollbacks: The LAPD discontinued its PredPol and LASER programs following audits and public criticism. Chicago ended its “Strategic Subject List” after independent evaluations found limited evidence of crime reduction. Santa Cruz banned predictive policing entirely, citing concerns that algorithms trained on historical data reinforce existing enforcement patterns rather than producing unbiased forecasts.[8]
    • Probable cause requirements: Several states now prohibit law enforcement from relying solely on facial recognition matches as the basis for arrests or probable cause, requiring independent evidence instead.[8] This approach addresses the root cause of documented wrongful arrests, primarily of Black individuals, by ensuring that an AI output triggers further investigation rather than direct action.

    The Judicial Perspective: Guardrails for the Courtroom

    As AI tools enter criminal courts, judges face questions about admissibility, transparency, and due process. A 2026 article published by the American Bar Association’s Criminal Justice Section and summarized by the National Association for Presiding Judges and Court Executive Officers outlines a framework for balancing innovation and constitutional protections.[9]

    The article identifies five guardrails that courts and agencies should adopt:

    • Transparency and validation: AI systems must be independently validated before use in criminal proceedings, and vendors must disclose training data sources, bias mitigation strategies, and technical materials for defense review.[9] Idaho’s 2019 transparency law, which prohibits vendors from invoking trade-secret defenses when their risk assessment tools are used in liberty decisions, is cited as the leading model.[9]
    • Human oversight and accountability: AI must be advisory, not determinative. Judges must retain interpretive authority and apply tools to facts rather than allowing tools to supplant judicial reasoning. Prosecutors and defenders must contextualize AI evidence and avoid automation bias.[9]
    • Continuous review: AI systems evolve over time, and accuracy may degrade as models are updated or datasets shift. Courts must demand periodic validation reports and require vendors to notify them of substantive model changes. If a system begins performing poorly or exacerbating disparities, it must be corrected, suspended, or withdrawn entirely.[9]

    The article emphasizes that the fundamental question is not whether AI should be used, as practice has already answered that, but how to use it responsibly. The authors warn that “embracing AI without scrutiny risks automating injustice” and conclude that “people, not algorithms, must remain the ones who make the decisions that determine liberty and justice.”[9]


    The Path Forward

    The current landscape presents a paradox. Federal AI deployment in criminal justice is accelerating rapidly, with 315 active use cases touching everything from inmate classification to predictive policing. At the same time, the federal government is actively trying to prevent states from building guardrails for the private sector’s development of these same technologies, even as state and local governments continue passing restrictions on their own agencies’ use.

    The executive order’s carve-out for state procurement and use of AI is significant because it preserves the most direct lever for criminal justice oversight. States can still regulate how their own law enforcement agencies purchase, deploy, and monitor AI systems. The question is whether they will use that authority aggressively enough to keep pace with the technology’s spread, or whether the chilling effect of federal threats will slow the guardrail-building process.

    The DOJ’s own December 2024 report provides a roadmap, with recommendations for community engagement before deployment, independent validation, data quality controls, and ongoing monitoring. Whether those recommendations survive the current administration’s deregulatory posture, and whether states continue filling the governance gap, will determine whether AI in criminal justice enhances fairness and efficiency or amplifies existing disparities under a veneer of algorithmic objectivity.


    References

    [1] [PRIMARY] U.S. Department of Justice. “AI Use Case Inventory 2025.” justice.gov/ai/ai-inventory. 2025. https://www.justice.gov/ai/ai-inventory

    [2] [SECONDARY] FedScoop. “DOJ ramps up AI for legal work, crime predictions, surveillance, inventory shows.” fedscoop.com. 2025. https://fedscoop.com/justice-department-artificial-intelligence-ai-surveillance-inventory-predictive-technology-algorithm-bias/

    [3] [PRIMARY] U.S. Department of Justice, Office of Legal Policy. “Artificial Intelligence and Criminal Justice, Final Report.” December 3, 2024. https://www.justice.gov/olp/media/1381796/dl

    [4] [SECONDARY] Council on Criminal Justice. “DOJ Report on AI in Criminal Justice: Key Takeaways.” counciloncj.org. April 2025. https://counciloncj.org/doj-report-on-ai-in-criminal-justice-key-takeaways/

    [5] [PRIMARY] The White House. “Ensuring a National Policy Framework for Artificial Intelligence, Executive Order 14365.” December 11, 2025. https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/

    [6] [SECONDARY] Bains, Chiraag; Chohlas-Wood, Alex; and Kinsey, Katie. “States can—and should—regulate AI in criminal justice.” Brookings Institution. April 16, 2026. https://www.brookings.edu/articles/states-can-and-should-regulate-ai-in-criminal-justice/

    [7] [SECONDARY] Kaste, Martin. “Live cameras are tracking faces in New Orleans. Who should control them?” NPR. December 16, 2025. https://www.npr.org/2025/12/16/nx-s1-5616681/new-orleans-live-facial-recognition-surveillance

    [8] [SECONDARY] Gao, Canyu. “Slowing Down Artificial Intelligence in Policing.” New Jersey State Policy Lab, Rutgers University. April 27, 2026. https://policylab.rutgers.edu/publication/slowing-down-artificial-intelligence-in-policing/

    [9] [SECONDARY] MacKenzie, Hon. Brian W. and Wallace, David, Esq. “AI in the Criminal Courts: Balancing Innovation and Justice.” American Bar Association Criminal Justice Section. Spring 2026. https://napco4courtleaders.org/2026/06/ai-in-the-criminal-courts-balancing-innovation-and-justice/


    AI automated and generated content. Human reviewed.


    Listen to the Bears Discuss This

    Episode 11 of AI Policy for the Bears

  • When the Computer Writes the Report: States Race to Regulate AI-Generated Police Narratives

    When the Computer Writes the Report: States Race to Regulate AI-Generated Police Narratives

    Utah and California have enacted the nation’s first laws governing generative AI in police report writing, exposing a transparency gap that products like Axon’s Draft One were deliberately built to widen.


    A New Frontier in Policing, Quietly Arriving

    Law enforcement agencies across the United States are adopting artificial intelligence tools faster than any legislature can write rules for them. The newest and most consequential application is also the most mundane: software that listens to body-worn camera audio and writes the police report for the officer. A June 2026 Stateline investigation found that companies including Axon, Motorola Solutions, TRULEO, Flock Safety, and Mark43 are selling AI products for surveillance, data analysis, and report writing to hundreds of agencies, with “no consistent national framework” governing their use.[1]

    The speed of adoption has alarmed legal scholars. Andrew Guthrie Ferguson, a George Washington University law professor and author of Your Data Will Be Used Against You, told Stateline that AI gives police and prosecutors “a lot more power over us in ways that I think will be deeply uncomfortable for many of us,” and warned of “very real constitutional, statutory and practical risks with this new model of agentic policing.”[1]

    What makes the report-writing application uniquely dangerous is that it operates inside the most important document in the criminal process. The police report is the narrative that prosecutors build charges on, that defense attorneys must rebut, and that judges weigh at trial. When a machine authors that narrative, the chain of accountability that the entire system depends on begins to blur.


    Axon’s Draft One: Built to Evade Audit

    The product driving most of the regulatory action is Draft One, made by Axon Enterprise, the company that dominates the body-camera market. Draft One uses a ChatGPT variant to process body-worn camera audio, generates report text with bracketed placeholders for officer observations, and then, by design, deletes everything. According to an Electronic Frontier Foundation investigation published in July 2025, Draft One does not save the first AI-generated draft or any subsequent versions, and there is no log indicating which portions of a report were AI-written versus officer-written.[2]

    Axon’s senior principal product manager for generative AI was blunt about the reasoning, telling a public records requester: “we don’t store the original draft and that’s by design and that’s really because the last thing we want to do is create more disclosure headaches for our customers and our attorney’s offices.” He added that once an officer copies the text into a third-party records system and closes the browser tab, “it’s gone. It’s actually never stored in the cloud at all.”[2]

    The consequence is that several questions fundamental to due process cannot be answered after the fact:

    • Are officers meaningfully editing the AI drafts, or rubber-stamping them? Without comparing the original draft to the final report, no one can tell whether an officer reviewed, corrected, or simply accepted the machine’s narrative.
    • When the AI makes an error, is there a pattern? Audio transcription can misread dialects, slang, and idioms. EFF notes that in a tense traffic stop, mistaking a metaphorical statement for a literal claim “could fundamentally change a police report’s content” and there is no record to check it against.[3]
    • If a report contains a fabrication, was it the officer or the machine? Officers who deliberately exaggerate gain “even more of a veneer of plausible deniability,” because they can claim the AI simply failed to capture what happened in chaotic footage.[3]

    The practical scale of the audit gap is striking. When the Lafayette Police Department in Indiana received a public records request for all Draft One reports, it responded: “We do not have the ability to create a list of reports created through Draft One. They are not searchable.” The Palm Beach County Sheriff’s Office could identify more than 3,000 Draft One reports only because it required officers to type the keyword “Draft One” into each one.[2]

    EFF also discovered that Axon engineers found a bug allowing officers to circumvent guardrails on at least three occasions, enabling submission of AI reports without reading them first. A records administrator at the Frederick Police Department in Colorado summed up the dynamic in an email: “We love having new toys until the public gets wind of them.”[2]


    Utah Moves First

    The first legislative response came in Utah. SB 180, signed into law and effective May 7, 2025, was the nation’s first statute addressing AI-generated police reports.[4] The law requires law enforcement agencies to adopt a written policy governing generative AI use, mandates a disclaimer on any report created wholly or partly by AI, and requires officers to certify that they checked the AI-generated content for accuracy.[3]

    Utah’s approach is deliberately minimal. It does not require agencies to retain the original AI draft, which means the core audit problem that EFF identified with Draft One, the inability to compare what the machine wrote against what the officer submitted, remains unsolved. EFF, which supported the bill, called it “a necessary step in the right direction” but “unfortunately” only a first step, arguing that a disclaimer alone “is not enough to curb potential harms.”[3]

    The logic of Utah’s disclosure-first model is that sunlight is the cheapest and fastest intervention. Forcing an agency to admit, on the face of a report, that a machine wrote part of it gives judges, defense counsel, and the public a fighting chance to scrutinize the document differently. But without the underlying draft, disclosure only tells you that AI was involved; it does not tell you what the AI actually produced before a human touched it.


    California Goes Further

    Five months later, California Governor Gavin Newsom signed SB 524, making California the second state to regulate AI in police report writing and, by most accounts, the most comprehensive. Effective January 1, 2026, the law, codified at California Penal Code section 13663, requires that when a report is generated wholly or partly by AI, the agency’s policy must ensure:[5][6]

    • Disclosure of the AI program used, on the report itself.
    • Officer certification that the officer reviewed the report for accuracy.
    • Retention of the first AI-generated draft for as long as the final report is retained, so that judges, defense attorneys, and auditors can see which portions were written by the officer versus the computer.
    • An audit trail tied to the source footage or audio.
    • A rule that only the final report counts as the officer’s statement, closing the loophole where an officer could blame the AI for inaccuracies.
    • A vendor data ban, prohibiting AI vendors from selling or sharing the information a police agency provided to the system.[5]

    The draft-retention requirement is the provision that matters most, and it is the one that Draft One cannot satisfy as currently designed. Because Axon deletes the original draft by design, EFF concluded that California departments using Draft One must either pressure Axon to change its product to include edit logs, have officers manually retain evidence of what their original draft looked like, or drop the tool altogether.[5]

    California’s innovation is to treat the AI draft as a form of evidence in its own right. The reasoning is that if the final report is going to send someone to prison, the system deserves to see what the machine produced before a human edited it, exactly as it would preserve any other piece of evidence that shaped an investigation. Retention turns a marketing feature (no copies floating around) into a legal liability.


    The Compliance Problem: Products Built to Defy the Law

    What the Utah and California laws reveal is a structural mismatch between how AI policing products are engineered and what the legal system requires. The same design choices that Axon markets as customer convenience, no stored drafts, no edit logs, no searchable index of AI-generated reports, are precisely the features that make meaningful oversight impossible.[2]

    As of mid-2026, the patchwork is still thin. According to a state-by-state tracker maintained by Barricade AI, only California and Utah have enacted laws specifically on AI-written police reports, though New York and Minnesota have pending bills and several prosecutors have imposed their own restrictions:[7]

    • New York (S10425 / A9253-A), the most ambitious pending bill, would require a public inventory of investigative AI systems, disclosure, first-draft retention, and enforcement by both the attorney general and private plaintiffs.[7]
    • Minnesota (SF 4575 / HF 4536) would go further than any enacted law by prohibiting generative AI in official records, including police reports, outright.[7]
    • Prosecutor-led moratoria have filled the statutory gap. The Connecticut Chief State’s Attorney imposed a statewide moratorium on AI-drafted criminal reports in April 2026. The King County, Washington Prosecuting Attorney’s Office announced in September 2024 that it would not accept AI-assisted report narratives, telling officers: “We do not fear advances in technology, but we do have legitimate concerns about some of the products on the market now.”[7]

    The prosecutorial pushback matters because prosecutors are the gatekeepers of what enters the courtroom. A police department can buy any tool it wants, but if the local district attorney refuses to file charges based on an AI-assisted report, the product has no practical value in that jurisdiction. That gives prosecutors leverage that state legislatures, which move slowly and face vendor lobbying, often lack.


    Beyond Reports: The Broader AI Policing Surge

    AI-written reports are the tip of a much larger deployment. The National Conference of State Legislatures reported in February 2025 that law enforcement uses AI in three ways: assisting humans with tasks, expanding human capabilities, and, in limited instances, replacing humans with fully automated processes, and that no state or locality had adopted a comprehensive set of laws governing any of them.[8]

    The applications NCSL catalogs span the investigative process: machine learning to predict criminal activity and tailor staffing; computer vision for facial recognition, fingerprint matching, and DNA analysis; natural language processing to draft reports and flag problematic officer interactions; autonomous drones for crash reconstruction and first response; and process automation for dispatch triage and court reminders.[8]

    Predictive policing, the use of algorithms to forecast where crime will occur and whom to police, remains the most contested application. The NAACP issued a policy brief calling on state legislators to regulate the practice, arguing that because the underlying data is drawn from “historical criminal data and police activity,” it is “inherently biased due to targeted over-policing and discriminatory criminal laws disproportionately impacting Black communities” (advocacy organization). The brief cites a letter from U.S. senators to the Department of Justice stating that “mounting evidence indicates that predictive policing technologies do not reduce crime, Instead, they worsen the unequal treatment of Americans of color by law enforcement.”[9]

    The NAACP’s recommendations reflect a core logic that runs through the whole debate: the problem is not the algorithm but the data and the accountability around it. The brief calls for independent oversight bodies, mandatory disclosure of data sources and methodologies, community engagement before deployment, a ban on biased historical data, and legal frameworks with penalties for civil liberty violations.[9]

    The misuse cases are already arriving. Stateline reported that in April 2026 a former Costa Mesa, California officer pleaded guilty to using law enforcement databases and Flock Safety cameras to monitor his wife, a mistress, and romantic rivals, with similar misuse allegations in eight other states. At least 30 cities have ended or canceled automated license plate reader contracts since early 2025 amid surveillance and data-sharing concerns.[1]


    A Framework for the Road Ahead

    The most detailed attempt to translate principle into practice comes from the Council on Criminal Justice, which in March 2026 published a user decision framework translating its October 2025 Principles for the Use of AI in Criminal Justice into operational steps for agencies evaluating AI tools.[10]

    The framework’s central insight is that procurement is the last and best safety net. Because criminal justice agencies cannot rely on vendor claims, contracts must establish enforceable performance standards, data rights, fairness requirements, auditability provisions, and termination rights. The framework requires agencies to screen every proposed system against a set of prohibited-use questions before deployment:[10]

    • Does the system make autonomous decisions about liberty without substantial human review?
    • Does it eliminate or impair the right to contest or appeal?
    • Does it circumvent constitutional protections like due process and equal protection?
    • Does it track or surveil lawful, constitutionally protected activity?
    • Does it target based on protected characteristics, creating unjustified discriminatory effects?
    • Does it systematically undermine human dignity?[10]

    A “yes” to any question triggers a requirement to mitigate through design changes, procedural safeguards, or alternative deployment; if mitigation is impossible, the system is prohibited. The framework also requires regular demographic impact assessments, mandatory training to counter automation bias, and human operators with clear authority to override AI recommendations.[10]

    What ties these threads together, the Utah and California statutes, the prosecutorial moratoria, the NAACP brief, and the CCJ framework, is a shared recognition that the default trajectory of AI in policing is opacity. The technology arrives faster than the rules, the vendors design for deniability, and the agencies adopt for efficiency. Every meaningful intervention, whether a disclosure requirement, a draft-retention mandate, a prosecutor’s memo, or a procurement checklist, is an attempt to force transparency back into a system that is being engineered to shed it. Whether the patchwork of state laws that emerged in 2025 and 2026 becomes a real constraint on vendors like Axon, or merely a disclosure label on an unaccountable process, will depend on whether enough jurisdictions follow California’s lead in demanding that the machine’s first draft survive.


    References

    [1] [SECONDARY] Watford, Amanda. “Police Use of Artificial Intelligence Grows as Rules Lag Behind.” Stateline. June 26, 2026. https://stateline.org/2026/06/26/police-use-of-artificial-intelligence-grows-as-rules-lag-behind/

    [2] [ADVOCACY] Guariglia, Matthew, and Dave Maass. “Axon’s Draft One Is Designed to Defy Transparency.” Electronic Frontier Foundation. July 10, 2025. https://www.eff.org/deeplinks/2025/07/axons-draft-one-designed-defy-transparency

    [3] [ADVOCACY] Guariglia, Matthew. “Utah Bill Aims to Make Officers Disclose AI-Written Police Reports.” Electronic Frontier Foundation. February 21, 2025. https://www.eff.org/deeplinks/2025/02/utah-bill-aims-make-officers-disclose-ai-written-police-reports

    [4] [PRIMARY] Utah State Legislature. “SB 180: Law Enforcement Usage of Artificial Intelligence.” 2025 General Session. https://le.utah.gov/Session/2025/bills/introduced/SB0180.pdf

    [5] [ADVOCACY] Guariglia, Matthew. “Victory! California Requires Transparency for AI Police Reports.” Electronic Frontier Foundation. October 14, 2025. https://www.eff.org/deeplinks/2025/10/victory-california-requires-transparency-ai-police-reports

    [6] [PRIMARY] California State Legislature. “SB 524: Law Enforcement Agencies: Artificial Intelligence.” 2025-2026 Session. https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202520260SB524

    [7] [SECONDARY] Barricade AI. “AI Police Report Laws by State: 2026 Tracker.” June 2026. https://barricade.tech/articles/ai-police-report-laws-by-state

    [8] [PRIMARY] Ezeh, Nicole, Amber Widgery, and Chelsea Canada. “Artificial Intelligence and Law Enforcement: The Federal and State Landscape.” National Conference of State Legislatures. February 3, 2025. https://www.ncsl.org/civil-and-criminal-justice/artificial-intelligence-and-law-enforcement-the-federal-and-state-landscape

    [9] [ADVOCACY] NAACP. “Artificial Intelligence in Predictive Policing Issue Brief.” https://naacp.org/resources/artificial-intelligence-predictive-policing-issue-brief

    [10] [SECONDARY] Rothman, Jesse. “Assessing AI for Criminal Justice: A User Decision Framework.” Council on Criminal Justice. March 2026. https://counciloncj.org/assessing-ai-for-criminal-justice-a-user-decision-framework/


    Listen to the Bears Podcast

    Episode 5: When the Computer Writes the Report — click to listen

    More episodes of AI Policy for the Bears →



    AI automated and generated content. Human reviewed.

  • The Federal AI Governance Gap: FBI Doubles Down While States Build Guardrails

    The Federal AI Governance Gap: FBI Doubles Down While States Build Guardrails

    As the FBI more than doubles its AI use cases, federal risk assessments remain incomplete — leaving states and cities to set the rules.


    The FBI’s AI Buildout

    In 2025, the Department of Justice released its annual AI inventory revealing that the FBI had more than doubled its AI use cases — from 19 in 2024 to 50 in 2025, with 27 categorized as law enforcement. The broader DOJ inventory listed 315 total AI entries with 188 active use cases, a 31 percent increase over the prior year. Law enforcement applications accounted for 62 percent of all deployed use cases.[1][2]

    Among the new deployments are approximately five AI projects for generating investigative leads using suggested facial matches and other biometric data, with four of the five already actively in operation. The FBI’s nine “high-impact” AI use cases include facial recognition, data mapping, biometric matching across large datasets, next-generation location tracking, and automatic vehicle identification.[1][2]

    The core logic of this expansion is straightforward: AI lets investigators search vastly larger datasets — billions of images, cross-referenced databases, real-time location feeds — far faster than any human team could. A single probe image can now generate dozens of candidate matches across jurisdictions in seconds. The promise is efficiency; the risk is that the speed and scale of AI-generated leads can overwhelm the traditional investigative checks that prevent wrongful arrests.

    Missing Risk Assessments

    Despite the rapid deployment, none of the FBI’s high-impact AI use cases have completed the required risk management assessments for issues like bias, safety, and performance. The Trump administration set an April 2026 deadline for all federal agencies to meet minimum risk management requirements for high-impact AI, but as of the inventory’s release, the work remained incomplete.[1][2]

    This matters because “high-impact” designations exist precisely for use cases where errors carry serious consequences — false arrests, misidentification, surveillance of innocent people. Without completed assessments, there is no documented evidence that the FBI has evaluated accuracy disparities across demographic groups, established performance thresholds, or put in place mitigation measures for known failure modes. The inventory itself is described by experts as vague, with tools “bunched together” and labeled generically as “search tool, chatbot, data outlier detection.”[1]

    “There’s a level of ambiguity in how they’ve chosen to describe some of the things in the document itself. It’s extremely difficult to get any kind of definitive sense in terms of how good it is, who’s conducting the oversight or auditing or stress testing.”

    — Patrick Eddingon, Senior Fellow, Cato Institute[1]


    Why Facial Recognition Produces Wrongful Arrests

    The danger of unmanaged AI in policing is not theoretical. At least seven people have been falsely arrested nationwide after facial recognition matches, and in every known case, police failed to conduct sufficient follow-up investigation before making the arrest.[3]

    The most documented case is that of Robert Williams, a Farmington Hills, Michigan resident arrested in January 2020 after facial recognition matched his expired driver’s license photo to surveillance footage of a robbery. Williams was the 9th-best match — not the top result — yet detectives focused on him without checking his alibi, collecting fingerprints, or interviewing witnesses present at the scene. The photo lineup was conducted by a store security employee who had not witnessed the crime. Williams spent over 30 hours in custody, handcuffed in front of his wife and two young daughters, before being released.[4]

    The underlying problem is structural. Facial recognition systems produce ranked lists of possible matches from databases that can contain billions of images. As databases grow, the pool of plausible-looking candidates expands, increasing the odds that an innocent person will appear as a suspect. Researcher Margaret Kovera of John Jay College of Criminal Justice explains that larger databases “up the number of innocent people who are appearing as suspects” while decreasing the ratio of correct identifications.[3]

    This problem is compounded by two factors:

    • Demographic bias: Facial recognition algorithms have historically performed worse on Black, Asian, and Indigenous faces. A 2017 NIST study of 140 algorithms found false positive rates up to 100 times higher for certain demographic groups, driven by training datasets built primarily on white faces and by digital cameras that fail to capture sufficient color contrast for darker-skinned individuals.[4]
    • Automation bias: Investigators exhibit a “hardwired tendency to believe computer outputs are correct, even when warned about errors.” When a system presents a confident-looking match, officers may shortcut the traditional investigative steps — witness interviews, physical evidence, alibi verification — that would catch a false positive.[3]

    As ACLU attorney Nate Wessler puts it, facial recognition is “a technology that is both dangerous when it works and dangerous when it doesn’t work.” When it works, it enables pervasive tracking that threatens constitutional freedoms. When it fails, it produces wrongful arrests of innocent people.[3]


    The Detroit Settlement: A Template for Guardrails

    In the summer of 2024, Williams reached a landmark settlement with the Detroit Police Department — the first in the nation where binding policy changes were negotiated as part of a facial recognition lawsuit. The settlement established what civil rights advocates call the strongest police department policies constraining facial recognition use in the country.[4][5]

    The key provisions include:

    • No arrest based solely on FRT results: Police cannot go directly from a facial recognition match to a warrant or witness identification without independent, reliable evidence linking the suspect to the crime.[4]
    • Lineup restrictions: Photo lineups cannot be based solely on an FRT lead, and the photo used in a lineup must be different from the one the facial recognition system returned.[5]
    • Mandatory training: Officers must receive training on FRT risks, dangers, and higher misidentification rates for people of color before using the technology.[4]
    • Defendant notification: Defendants must be informed when facial recognition was used in their case, including details about the system and process.[5]
    • Retroactive audit: Detroit Police must audit all cases since 2017 in which facial recognition was used to obtain an arrest warrant.[4]

    These measures go significantly beyond the “no sole basis for arrest” rule adopted in seven states. The settlement’s logic is that a blanket prohibition on FRT-only arrests, while necessary, is insufficient — it still permits facial recognition to serve as the overwhelming basis for an arrest, creating confirmation bias downstream. By requiring independent corroboration, restricting lineup procedures, and mandating disclosure to defendants, the settlement addresses the full investigative chain where errors compound.[5]


    States Fill the Federal Vacuum

    While Congress has introduced facial recognition legislation but not passed any, states have been far more proactive. As of the end of 2024, fifteen states have laws limiting police use of facial recognition, up from just one in 2017. These laws span states across the political spectrum, from Montana and Utah to Maryland and Massachusetts.[5]

    The state approaches fall into three tiers of increasing strength:

    • Narrow limits (2 states): New Hampshire and Oregon prohibit facial recognition on body-worn cameras only — the narrowest restriction, addressing the most immediate real-time surveillance concern without limiting broader investigative use.[5]
    • One stronger limit (6 states): Alabama, Illinois, Minnesota, Massachusetts, New Jersey, and Vermont each adopted at least one of three key guardrails: a warrant requirement, mandatory defendant notification, or a serious-crime-only restriction. New Jersey’s appellate court ruled in New Jersey v. Arteaga (2023) that defendants must be notified of facial recognition use to protect due process rights under Brady v. Maryland.[5]
    • Multiple stronger limits (7 states): Colorado, Maryland, Maine, Montana, Utah, Virginia, and Washington combine two or more guardrails. Montana became the first state to require a warrant for police facial recognition use (2023), and Utah became the second (2024).[5]

    The warrant requirement is emerging as the foundational policy recommendation. Jake Laperruque of the Center for Democracy and Technology argues it “simultaneously is a strong shield against abuse and should have little disruptive effect on legitimate law enforcement needs” — because the primary legitimate use of facial recognition (identifying an unknown perpetrator from an image of a crime) already provides the probable cause a warrant requires. What a warrant requirement blocks are the nefarious uses: cataloging protester identities, tracking worship attendees, surveilling communities without individualized suspicion.[5]


    New York City Strengthens Oversight

    At the municipal level, New York City passed an expanded POST Act legislative package in April 2025, strengthening oversight of NYPD surveillance technology. The original 2020 Public Oversight of Surveillance Technology Act aimed to bring accountability to NYPD surveillance tools, but Department of Investigation audits revealed persistent compliance gaps, noncompliance, and use of surveillance tools without full public disclosure.[6]

    The expanded package includes three key bills:

    • Intro 168 (Farías): Empowers the DOI to request detailed, itemized lists of all NYPD surveillance technologies, with biannual reporting on technologies acquired or retired, including access and retention policies.[6]
    • Intro 233 (Hudson): Mandates the NYPD publish a publicly accessible facial recognition use policy, with biannual audits of facial recognition use reported to the DOI and posted online.[6]
    • Intro 480 (Won): Expands surveillance Impact and Use Policies to require disclosure of all outside entities receiving NYPD surveillance data, safeguards for each entity, reporting on physical safety risks and disparate impacts on protected communities, and an internal tracking system for intergovernmental data sharing.[6]

    In response, the NYPD revised its Facial Recognition Impact and Use Policy in February 2026 to comply with Local Law 56 of 2025. The revised policy removed the prior claim that facial recognition software “does not use AI/ML” (noting the lack of industry-standard definitions), expanded safeguards around information security and access removal, and clarified that the technology generates “possible match candidates” treated as investigative leads only — not probable cause for arrest.[7]

    The NYPD policy explicitly states that facial recognition results “does not by itself establish a basis for a stop, probable cause to arrest, or to obtain a search warrant,” and requires peer review, supervisor approval, and manual visual comparison of all candidate matches before a possible match report is issued. Notably, the policy prohibits real-time facial recognition on CCTV, drones, and body-worn cameras.[7]


    The Federal-State Divergence

    The contrast between federal and sub-federal governance of AI in law enforcement is stark. The FBI is deploying facial recognition, biometric matching, and location tracking with no completed risk assessments and vague public reporting. Meanwhile, states are enacting warrant requirements, cities are mandating audits and public disclosure, and civil society coalitions of over 40 organizations are calling for outright moratoria on law enforcement face recognition.[1][8]

    The Brennan Center for Justice and its coalition partners argue that “the unfettered use of face recognition technology by law enforcement poses urgent threats to democracy” — because when anyone can be identified at any time, constitutional protections for assembly and association become illusory. Their coalition identifies six core civil rights concerns: exacerbation of policing harms in marginalized communities, privacy threats from persistent tracking, chilling of First Amendment activity, due process violations, non-consensual image use, and racial bias in misidentification.[8]

    DOJ’s own transparency falls behind even other federal agencies. Experts note the DOJ inventory is worse than DHS’s — “a traditionally tight-lipped organization” — with three clearly labeled facial recognition instances all showing 2025 deployment dates despite reports the FBI has used the technology for years. This is the second consecutive year of such reporting shortcomings.[1]

    What Comes Next

    The April 2026 federal risk management deadline will be a critical test. If the FBI and DOJ complete meaningful impact assessments — evaluating bias, establishing accuracy thresholds, and documenting mitigation measures — it would represent a significant step toward parity with state-level guardrails. If the deadline passes with assessments still incomplete or superficial, the federal-state governance gap will widen further, and the real rules governing AI in American policing will continue to be written in state capitals and city halls rather than in Washington.[1][2]

    The trajectory of state and local action suggests the direction of travel. Seven additional states are considering new facial recognition legislation, and the Policing Project has developed a bipartisan legislative checklist recommending transparency standards, officer training, procedural limits, defendant disclosure, vendor documentation, and enforcement mechanisms. The question is whether federal agencies will adopt these same principles voluntarily — or whether Congress will eventually force them to.[3][5]


    References

    [1] FedScoop — “FBI expands AI-powered biometric, facial recognition capabilities, inventory shows” (February 2025)

    [2] FEDagent — “DOJ Expands AI Use for Law Enforcement, FBI Doubles AI Applications in 2025” (February 5, 2025)

    [3] Stateline (Pew) — Gross, Paige. “Facial recognition in policing is getting state-by-state guardrails.” February 4, 2025. https://stateline.org/2025/02/04/facial-recognition-in-policing-is-getting-state-by-state-guardrails/

    [4] University of Michigan Law Quadrangle — Morioka, Sharon. “Flawed Facial Recognition Technology Leads to Wrongful Arrest and Historic Settlement.” Winter 2024-2025. https://quadrangle.michigan.law.umich.edu/issues/winter-2024-2025/flawed-facial-recognition-technology-leads-wrongful-arrest-and-historic

    [5] TechPolicy.Press — Laperruque, Jake. “Status of State Laws on Facial Recognition Surveillance: Continued Progress and Smart Innovations.” January 6, 2025. https://techpolicy.press/status-of-state-laws-on-facial-recognition-surveillance-continued-progress-and-smart-innovations

    [6] NYC Council — “City Council Passes Expanded POST Act Legislative Package to Strengthen Transparency and Oversight of NYPD Surveillance Technology.” April 15, 2025. https://council.nyc.gov/amanda-farias/2025/04/15/city-council-passes-expanded-post-act-legislative-package-to-strengthen-transparency-and-oversight-of-nypd-surveillance-technology/

    [7] NYPD — “Facial Recognition: Impact and Use Policy.” Revised February 4, 2026. https://www.nyc.gov/assets/nypd/downloads/pdf/public_information/post-final/Facial-Recognition-NYPD-Impact-and-Use-Policy-2.4.26-FINAL.pdf

    [8] Brennan Center for Justice — “Coalition Statement Highlights Major Civil Rights Concerns with Face Recognition.” June 3, 2021. https://www.brennancenter.org/our-work/research-reports/coalition-statement-highlights-major-civil-rights-concerns-face


    AI automated and generated content. Human reviewed.

  • Fifteen States and Counting: The Movement to Regulate Facial Recognition in Law Enforcement

    Fifteen States and Counting: The Movement to Regulate Facial Recognition in Law Enforcement

    How state-level guardrails are emerging as Congress stalls — and why warrant requirements are becoming the new standard.


    The Problem: AI Already Deployed, Rules Catching Up

    Facial recognition is not a future technology in American policing. It is already broadly deployed, and real people have already been harmed by it. In Colorado, police held a woman and her four children — including a six-year-old — at gunpoint based on a false automated license plate reader match [1]. In Georgia, a Black man was arrested and jailed for nearly a week for a retail theft in Louisiana, a state he had never visited, because of an erroneous facial recognition match [1]. Multiple wrongful arrests — disproportionately of Black individuals — have been documented across the country [1].

    The core problem is that these tools are being used without independent testing, judicial oversight, or accuracy standards. As the Brookings Institution notes, “most new AI technologies pitched to the criminal justice system remain untested by credible, independent sources, and increasingly lofty marketing claims are unlikely to hold up under scrutiny” [1].


    The State Response: Fifteen States Step In

    With Congress failing to pass federal legislation, states have become the laboratories of democracy for facial recognition policy. As of early 2025, fifteen states now limit police use of facial recognition technology — up from just one state in 2017 [2]. These states span a broad political spectrum, from Massachusetts to Montana.

    The safeguards fall into three tiers of increasing strength:

    • Body camera ban only (weakest): States like New Hampshire and Oregon simply prohibit facial recognition on police body cameras. This is a minimal floor that leaves most uses unregulated [2].
    • Single stronger limit: States like Illinois, Minnesota, and New Jersey have enacted one meaningful safeguard — such as a warrant requirement, defendant notification requirement, or a restriction to serious crimes only [2].
    • Multiple stronger limits (strongest): Colorado, Maryland, Maine, Montana, Utah, Virginia, and Washington combine several safeguards, creating layered protections that address different failure modes [2].

    The logic here is that no single safeguard is sufficient. A warrant requirement without a serious-crime limit could still allow police to use facial recognition for minor infractions. A notification requirement without accuracy testing could still produce wrongful arrests from faulty matches. The strongest state models stack protections the way a defense-in-depth strategy works in cybersecurity.


    The Warrant Requirement: Why It Matters

    The most consequential trend is the emergence of warrant requirements. Montana became the first state to require a warrant for police use of facial recognition in 2023, followed by Utah in 2024 [2]. The Center for Democracy and Technology argues that “a warrant requirement should become the foundation for facial recognition policy going forward” [2].

    The reasoning is straightforward: the primary legitimate use of facial recognition — identifying an unknown perpetrator captured in an image of a crime — inherently provides probable cause. The photo itself is evidence that a crime occurred. A warrant requirement blocks the illegitimate uses: cataloging protesters’ identities at a demonstration, tracking who enters a house of worship, or monitoring people going about their daily lives [2].

    Utah’s law, however, reveals the dangers of poorly drafted exceptions. It exempts law enforcement buildings, critical infrastructure, courthouses, public schools, and airports — and allows surveillance based on “reasonable articulable suspicion of a threat to commit a violent felony” without judicial review [2]. The concern is that this exception could permit untargeted crowd scanning, effectively creating a loophole that swallows the rule.


    Predictive Policing: The Broader Constitutional Problem

    Facial recognition is the visible front of AI in law enforcement, but predictive policing algorithms present a deeper constitutional challenge. These systems feed historical crime data, geographic information, and demographic details into machine learning models that assign “risk scores” to areas or individuals [3]. The fundamental problem is that the data driving these tools comes from historical police activity — which is inherently biased due to decades of targeted over-policing in minority communities [4].

    This creates a self-perpetuating cycle: biased data leads to biased predictions, which lead to increased police presence in minority neighborhoods, which generates more arrest data, which reinforces the bias in the next iteration. The NAACP warns that these tools “worsen the unequal treatment of Americans of color by law enforcement” rather than reducing crime [4]. U.S. Senators have called on the DOJ to cease funding predictive policing systems pending audits [4].

    The constitutional dimensions are significant. Under the Fourteenth Amendment’s Equal Protection Clause, proving a violation traditionally requires demonstrating discriminatory intent, not just discriminatory impact [3]. Since algorithms are designed to be objective — they lack human intent — they can produce deeply discriminatory outcomes without triggering constitutional scrutiny. Advocates argue for a “disparate outcome” standard, similar to employment law under Title VII, which would address bias without needing to prove that anyone intended it [3].

    The Fourth Amendment adds another layer. In Carpenter v. United States, the Supreme Court ruled that accessing historical cell phone location data without a warrant violated the Fourth Amendment [3]. Predictive policing similarly creates risk profiles based on data patterns rather than concrete suspicion of wrongdoing, raising the question of whether algorithmic surveillance requires its own warrant standard.


    The Robert Williams Case: What Happens When It Goes Wrong

    The most instructive case study is Robert Williams, the first known person wrongfully arrested due to facial recognition error. In 2024, Williams reached a settlement with Detroit police that established new guardrails for how facial recognition can be used [2]:

    • Police must receive training on risks and dangers before using facial recognition
    • Defendants must be notified when facial recognition was used, with details
    • Corroborating evidence is required before placing an individual in a photo lineup
    • Police must use a different photo than the one returned by the facial recognition system
    • Police must not tell witnesses reviewing lineups that a facial recognition match was found

    The last point is particularly important. Seven states have “not sole basis” rules prohibiting arrests based solely on a facial recognition match [2]. But in practice, police can still use a match as the primary basis for investigation, then ask a witness to confirm a single photo — creating confirmation bias that taints the entire process. The Williams settlement addresses this by requiring a different photo in the lineup and prohibiting officers from telling witnesses about the AI match [2].


    The Federal Obstacle: Trump’s Executive Order

    In December 2025, the Trump administration issued an executive order titled “Eliminating State Law Obstruction of National Artificial Intelligence Policy,” which attempts to chill state-level AI regulation through DOJ lawsuits, funding threats, and FTC preemption declarations [1]. However, the Brookings Institution argues that “the administration’s threats are built on legal quicksand” [1].

    Critically, the order specifically exempts laws regulating “state government procurement and use of AI” — meaning states remain free to regulate their own law enforcement agencies’ AI use [1]. This carveout is significant because most facial recognition regulations target exactly that: the rules governing how state and local police use these tools.


    What Still Needs to Happen

    • Move beyond “sole basis” rules: Seven states prohibit facial recognition matches as the “sole basis” for arrest, but this standard is too weak. The Williams settlement shows that corroborating evidence requirements and lineup safeguards are needed to prevent confirmation bias [2].
    • Close exception loopholes: Utah’s warrant requirement includes broad exceptions for schools, airports, and “reasonable articulable suspicion” — standards that could permit untargeted surveillance without judicial review [2]. Future legislation must narrowly tailor exceptions.
    • Mandate accuracy testing: Only Colorado and Virginia require testing and accuracy standards for facial recognition systems [2]. Without independent validation, vendors’ marketing claims go unchecked.
    • Address predictive policing: Facial recognition regulation is necessary but insufficient. Predictive policing algorithms present distinct constitutional challenges under both the Fourth and Fourteenth Amendments that require their own regulatory framework [3].
    • Ban biased data sources: The NAACP recommends prohibiting the use of historical crime data in predictive algorithms, since this data encodes decades of discriminatory policing [4]. This is the only way to break the feedback loop.
    • Require transparency: Proprietary algorithms prevent public scrutiny. Agencies should be required to disclose data sources, methodologies, and impact assessments [4].

    References

    [2] Laperruque, Jake. “Status of State Laws on Facial Recognition Surveillance: Continued Progress and Smart Innovations.” TechPolicy.Press. January 6, 2025.

    [1] Bain, Chiraag, Alex Chohlas-Wood, and Katie Kinsey. “States Can—and Should—Regulate AI in Criminal Justice.” Brookings Institution. April 16, 2026.

    [4] NAACP. “Artificial Intelligence in Predictive Policing: Issue Brief.”

    [3] Lin, Leo. “Algorithmic Justice or Bias: Legal Implications of Predictive Policing Algorithms in Criminal Justice.” Johns Hopkins University Law Review. January 1, 2025.


    AI automated and generated content. Human reviewed.