Tag: artificial intelligence

  • The Doctor’s Shield: How a No-Licensure-for-AI Vote, a Six-Demand Vendor Liability Push, and an 81-Percent Adoption Rate Left Vendors Exposed and Courts Unpersuaded

    The Doctor’s Shield: How a No-Licensure-for-AI Vote, a Six-Demand Vendor Liability Push, and an 81-Percent Adoption Rate Left Vendors Exposed and Courts Unpersuaded

    On November 9, the American Medical Association’s House of Delegates voted that AI cannot be licensed as a physician, then adopted a six-part plan to move liability for defective clinical AI onto vendors and any organization that mandates their tools. With four in five physicians already using AI and a quarter reporting no training on it, the association is trying to write the liability rules before the courts do.


    A Line Drawn in Orlando

    At its Interim Meeting in Orlando on November 9, 2026, the American Medical Association (AMA) House of Delegates adopted policy drawing what the association called “a clear line between artificial intelligence as a tool to support healthcare and the practice of medicine”: AI can augment physicians and assist with clinical tasks, “but it cannot be granted physician licensure, professional titles or certification to practice medicine” [1]. The policy calls for AI-enabled technologies to be integrated into physician-led teams and used at the direction of the treating physician, and for the AMA to advocate model legislation that distinguishes the authorization of a medical technology from professional licensure while prohibiting nonhuman entities from holding titles reserved for physicians [1].

    “AI can help physicians analyze information, identify patterns and support clinical decision-making, but patients deserve to know who is responsible for their care,” said AMA President Willie Underwood III, MD, MSc, MPH. “Physicians also need the authority, oversight and liability protections to make independent decisions when AI is part of the care team.” [1]

    The distinction the AMA wants legislated is between a product’s clearance and a professional’s license. FDA authorization says a tool may be marketed; licensure says a person is accountable for judgment. The core logic of the AMA’s position is that blurring the two lets vendors borrow the authority of medicine without inheriting its obligations — patients hear “doctor,” but no licensed professional stands behind the output. A companion directive would also bar AI developers from using patients’ clinical data for proprietary or research purposes without explicit authorization [2].

    Why Now: Adoption Is Outrunning the Rules

    The push is not happening in a vacuum. The AMA’s 2026 Physician Survey on Augmented Intelligence found that 81 percent of physicians now use AI professionally — more than double the 38 percent rate in 2023 — with the average physician reporting 2.3 distinct use cases, up from 1.1 three years ago [3]. The most common uses are medical research summarization and clinical care documentation, and more than three-quarters of physicians now believe AI improves their ability to care for patients [3].

    But the same survey shows the binding constraint. Clear liability frameworks “rank highest among regulatory actions essential to build physician trust and increase adoption of AI tools,” the AMA reported, alongside physician demands for data privacy (86 percent) and validated safety and efficacy (88 percent) [3]. Trade coverage of the survey data found roughly a third of physicians citing “clear liability frameworks” as the primary driver of adoption, and a quarter of physicians reporting they have received no AI training at all while 92 percent want more [4]. Eighty-eight percent of physicians told the AMA they are concerned about skill loss, particularly those with ten or fewer years in practice [3].

    The AMA’s framing of the problem is that physicians are being asked to absorb risks the technology creates. Physicians are currently “being asked to become a safety net to protect against clinical AI failures,” meaning they “may be held accountable for risks that they cannot control and, ‘in some cases, may not even be able to identify,’” said Jennifer Goldsack, head of the AMA Center on Digital Health and AI [4]. The underlying mismatch: adoption is compounding while accountability rules are undefined — which is precisely the vacuum the new policy is designed to fill.

    The Six Demands: Moving Liability Upstream

    The operative fact behind the liability plan is scale. The FDA has authorized more than 1,450 AI-enabled medical devices as of December 2025, and the AMA argues that — “no different from say, from an inherent problem in an MRI machine” — physicians should not bear responsibility for defects in products they did not build, cannot inspect, and often cannot see through [2]. To that end, the House of Delegates directed the AMA to advocate six legislative demands [2]:

    • Vendor liability for erroneous outputs. Federal and state legislation establishing that when a clinical AI product produces an erroneous output, and the physician used the product as intended while exercising appropriate clinical judgment, liability for patient harm rests with the AI vendor, manufacturer, or distributor — not a medical malpractice claim against the physician.
    • Device-grade product liability. AI vendors held to the same product liability standards as manufacturers of any other medical device, including liability for design defects and for failing to provide adequate instructions or warnings.
    • Insurance as a condition of marketing. Legislation requiring AI vendors and manufacturers to maintain product liability insurance adequate to cover claims arising from AI failures in clinical settings, tied to FDA marketing authorization.
    • A ban on liability-shifting contracts. Legislation prohibiting contractual provisions — including end-user license agreements and vendor contracts — that attempt to shift liability for AI tool failures from the manufacturer to the physician or health system.
    • Mandatory performance disclosure. Federal legislation requiring vendors to disclose known failure rates, limitations, and performance data for clinical AI products, including performance across different patient populations, so physicians can make informed decisions about the outputs they act on.
    • The mandated-tool rule. When a health system, hospital, payer, or other entity mandates the use of a specific AI tool and that tool produces an error causing patient harm, the mandating entity and the vendor bear the liability — the physician who was required to use the tool and exercised appropriate clinical judgment is not liable [2][5].

    Read together, the six demands share one allocation principle: liability follows control. Insurance requirements force vendors to price their own failure rates into their products; disclosure duties give physicians the information they would need to exercise the judgment the law expects of them; the anti-shifting rule prevents procurement contracts from quietly undoing what statute assigns; and the mandated-tool rule recognizes that when a system mandates a tool, the mandate — not the clinician holding it — is the decision that put the product in the room. The demands also extend to the machinery that judges physicians: the delegates called for physician-led oversight of any AI used in credentialing, privileging, peer review, and quality review, with full disclosure of the tool’s data sources, validation, limitations, and intended use [2].

    The Framework Underneath: Duty Attached to Control

    The November policy builds on a framework the AMA released in October 2026, in an issue brief titled “Liability for Clinical AI.” Its organizing principle is that “responsibility should reflect each party’s authority, access to information and ability to prevent or mitigate harm” [6]. Under that principle, four parties hold distinct duties: physicians are responsible for the clinical judgment they can meaningfully exercise; developers for the systems they build and the risks within their control; deployers — hospitals, health systems, and practices — for where and how AI is put to work; and patients must retain meaningful avenues for accountability and appropriate remedies when harm occurs [6].

    Goldsack’s illustration of the principle is a case every clinician will recognize: if a physician builds a treatment plan on the recommendation of an AI clinical decision support tool, but the tool failed to account for critical data and the physician had no visibility into that error, the physician should not be held responsible for the resulting harm [4]. Disclaimers, in the AMA’s view, are “not strong enough safeguards” — the framework instead asks where the risk of harm lies, which stakeholder can prevent it, and what information that stakeholder needs [4]. “This [framework] is going to create the rubric by which every different stakeholder gets very clear on their responsibility to proactively implement best practices where they have the ability to mitigate risk,” Goldsack said [4].

    The reason a framework is needed at all is structural. Traditional malpractice law assumes one licensed human holds all the relevant control over a patient’s care, which is what makes a single-defendant lawsuit coherent. When control fragments — a vendor designs the model, a hospital deploys it, a physician applies it — the old model either overcharges the physician, who lacks control over the defect, or under-deters the vendor, whom the injured patient cannot even see. The AMA’s answer is to re-attach legal duty to actual control, so that accountability and authority stay aligned as AI systems take on more autonomous clinical functions [6].

    What the Malpractice Bar Says

    The trial lawyers who will actually litigate these cases are not waiting for the AMA’s framework, and they see the problem differently. Medical malpractice attorney Rick Groffsky of Sommers Schwartz argues the technology is new but the legal principles are not: courts still expect clinicians to exercise independent medical judgment. “An AI tool is one more input to consider in that judgment, but it is not a replacement for it,” Groffsky said. “If a physician relies on a flawed recommendation without applying independent clinical judgment, the physician and the employing hospital will likely bear the responsibility” [7].

    The documentation AI generates cuts both ways. “The defense that ‘the computer told me to’ does not hold up, because the law still expects independent judgment,” Groffsky noted — but when a clinician overrides a correct AI warning, “the documentation now works against them,” because there is a record that the system flagged the risk [7]. He expects future cases to involve multiple defendants, since a physician may rely too heavily on AI while “the hospital may have handed that physician an AI tool it never adequately vetted or properly trained staff to use” — and he expects vendors to face direct claims where “the flaw was baked into the product itself,” as when training data underrepresents particular patient populations and the model consistently misses conditions affecting those groups [7].

    The courts are already arriving. A Florida lawsuit alleging that artificial intelligence contributed to a patient’s delayed cancer diagnosis is drawing fresh attention to the allocation question as AI moves deeper into diagnosis [7]. And Groffsky identifies what he calls perhaps the greatest long-term concern: automation bias. When software performs accurately most of the time, clinicians naturally become less skeptical, and the rare failures can be catastrophic. “The organizations most at risk are the ones that have deployed these tools widely but have not given clinicians clear guidance on when to rely on them and when to question them,” he said [7].

    Where the AMA and the Courts Will Collide

    The hinge in every one of the AMA’s demands is the phrase “appropriate clinical judgment.” The association wants physicians protected when errors are invisible to them; the malpractice bar observes that the electronic record now shows exactly what the tool recommended, what the clinician decided, and where the two diverged [4][7]. The collision point is the AMA’s demand for protection “against liability for choosing not to use an AI tool or follow an AI recommendation” [1] — a protection Groffsky’s analysis suggests courts will not extend to a clinician who dismisses a validated warning without documented reasoning, because both overreliance and underreliance on AI are expected to generate claims [7]. As certain diagnostic tools approach something resembling the standard of care, organizations choosing not to adopt them may face scrutiny comparable to ignoring any other established clinical practice [7].

    There is also a second-order tension the policy exposes. The AMA wants liability to follow control — but physicians are the one party in the chain whose control cannot be measured directly; it is inferred from training, documentation, and behavior. Vendors’ control is embedded in the product, and a hospital’s control is embedded in its procurement and training decisions. Physician control lives only in the judgment call itself, which is why the AMA pairs its shield with information-forcing demands: the disclosure and training requirements are what make “appropriate clinical judgment” legally defensible at all [3][2]. “When we get this right,” Goldsack argued, patients harmed by AI “can seek compensation from the appropriate party,” and harm becomes less likely in the first place “because every stakeholder understands what they’re responsible for” [4].

    What to Watch

    For hospitals and health systems

    The mandated-tool rule is the demand with the most immediate operational bite. Any system that requires clinicians to use a specific AI product would share liability with the vendor when that product fails — which means mandate decisions, and the documentation behind them, just became legal exposure decisions [2][5]. Systems should also expect pressure on the training gap, and should treat Groffsky’s advice as the operative standard: “Clinicians need real training on the tool’s limitations and failure modes, not just a software tutorial,” with governance and validation records built before deployment, not after [7].

    For AI vendors

    Every demand transfers cost onto the vendor side of the table: insurance requirements, failure-rate disclosure across patient populations, device-grade product liability, and a ban on the contract clauses that today shift those risks back to hospitals [2]. Provider groups including the AMA have separately told the FDA that developers — not providers — should be responsible for monitoring generative AI-enabled medical devices after they reach the market [4].

    For policymakers

    Watch state legislatures for the AMA’s model bills on licensure boundaries and vendor liability, and watch malpractice dockets for the cases that will define “independent judgment” in an AI-assisted exam room. The association’s stated goal is that clinical authority and accountability remain aligned as AI assumes increasingly autonomous functions [6] — but between now and then, the practical rule of thumb for every party in the chain is unchanged: control is the liability question, and the party that cannot document its control should not expect to keep its shield [7].


    References

    [1] American Medical Association — [ADVOCACY] “AMA: Artificial intelligence can augment physicians but don’t call it doctor.” Press release, November 10, 2026. (physician professional association)

    [2] American Medical Association — [ADVOCACY] “Who can be licensed as a physician? Not AI.” Policy adopted at the 2026 AMA Interim Meeting, November 10, 2026. (physician professional association)

    [3] American Medical Association — [ADVOCACY] “AMA: AI usage among doctors doubles as confidence in technology grows.” 2026 Physician Survey on Augmented Intelligence, 2026. (physician professional association)

    [4] Vaidya, Anuja — [SECONDARY] “AI vendors, health systems join physicians in liability for clinical AI: AMA.” Healthcare Dive, November 2, 2026.

    [5] Bruce, Giles — [SECONDARY] “AMA says AI can’t be licensed as a physician, pushes vendor liability.” Becker’s Hospital Review, November 10, 2026.

    [6] American Medical Association, Center on Digital Health and AI — [ADVOCACY] “Liability for Clinical AI.” Issue brief, October 2026. (physician professional association)

    [7] Siwicki, Bill — [SECONDARY] “AI misdiagnoses raise new liability questions for health systems.” Healthcare IT News, July 27, 2026.


    AI automated and generated content. Human reviewed.

  • The Provenance Turn: How a Deleted Dataset, an Everyone-in-the-Group Order, and a Fee Clock Running Out This Week Moved the AI Copyright Fight From Training to Acquisition

    The Provenance Turn: How a Deleted Dataset, an Everyone-in-the-Group Order, and a Fee Clock Running Out This Week Moved the AI Copyright Fight From Training to Acquisition

    Fair use survived summary judgment in Kadrey v. Meta — and barely mattered. The live questions in the five Meta copyright cases now sharing a May 3, 2027 trial date are how roughly 196,000 pirated books reached Meta’s servers, who inside the company approved the torrenting, and whether the command histories that would prove it still exist.


    The Question That Survived Summary Judgment

    When Judge Vince Chhabria ruled for Meta on summary judgment in Kadrey v. Meta Platforms on June 25, 2025, he held that training the Llama models on copyrighted books was fair use on the record before him — two days after Judge William Alsup reached the same conclusion for Anthropic’s models in Bartz v. Anthropic [1]. But the win came with a flashing caveat. “This ruling does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful,” Chhabria wrote. “It stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.” [1]

    The argument the plaintiffs failed to develop was market dilution: that Meta copied their works to build a product likely to flood the market with substitutes. Chhabria called it the “potentially winning argument” and observed that LLMs are a technology courts have never confronted — simultaneously so transformative that copying for training favors fair use, and so capable of mass-producing substitute works that it can gut the market for the originals. “In cases involving uses like Meta’s, it seems like the plaintiffs will often win,” he wrote, “at least where those cases have better-developed records on the market effects of the defendant’s use.” [1]

    What the ruling did not resolve was acquisition. The court allowed the authors to keep pursuing claims that Meta distributed their works through BitTorrent — the downloads themselves, and the uploads the protocol performs as the price of downloading [2]. A contributory infringement claim based on the same torrenting was added on April 1, 2026 [3]. Sixteen months after the fair-use ruling, that residual claim has become the whole case: five actions against Meta are managed together in the Northern District of California, and nearly every contested issue on the docket concerns how the pirated books got in — not what happened to them once they arrived [4][5].


    The Deposition That Contradicted an Interrogatory

    In a joint case management statement filed October 27, 2026, the rightsholders reported deposition testimony from Stephen Roller, a researcher who worked at Meta from 2017 to 2023: he torrented and knowingly distributed Books3 — the dataset of roughly 196,000 pirated books scraped from the private tracker Bibliotik — during the years he worked at the company, and Meta had copies of the dataset by the time he left its AI research group FAIR [5]. Roller no longer has his own copy. The plaintiffs say he deleted it about a year ago “due to all this litigation,” as it “seem[ed] contentious to have copies of The Pile” — the composite dataset that contains Books3 [4].

    The testimony cuts against what Meta had said under oath. As recently as its August 24, 2026 interrogatory responses, Meta stated that it did not believe it obtained Books3 via BitTorrent or any other peer-to-peer protocol [5].

    Meta’s half of the statement disputes the framing. Roller “did not know when he downloaded a copy of the ‘The Pile’ via torrent, and did not testify that his download of the ‘The Pile’ was for Meta or at Meta, noting it could have been for himself or for another employer,” the company wrote — though Roller acknowledged that Books3 was created in 2020, while he worked at Meta, and that he worked for no other company during those years [4]. Meta adds that the deleted copy sat on Roller’s personal computer and was removed more than two years after he left, because he “didn’t need it, so [he’d] rather not have it” [4]. What is not in dispute is that Meta has acknowledged since early 2024 that it used portions of Books3 to train Llama 1 and Llama 2 — without ever saying how it obtained the dataset [4].

    The core logic of the fight is structural. BitTorrent does not separate taking from giving: a client downloads by simultaneously uploading pieces to other users. So even with training on the books held to be fair use, the protocol itself creates a distribution claim — a separate exclusive right, with separate exposure — that survives the training defense entirely. That is why the deposition of a mid-level researcher about a copy on his personal machine is now the center of the case. [4][5]


    “Everyone in the Group”: The Order That Redefined Decision-Maker

    In October 2026, Magistrate Judge Thomas Hixson resolved a fight over an Entrepreneur Media interrogatory that asked Meta to identify everyone who “made, authorized, or approved” the decisions to obtain shadow-library datasets, and to do so via BitTorrent. Meta’s answers named personnel involved in data decisions but said no one “formally” made or approved the choices — they were often the result of team consensus, and the torrenting itself was described as “technical decisions” by the engineers who ran the downloads. The company added that the 2024 acquisition from Anna’s Archive went through an internal legal review involving its in-house lawyers [6].

    Entrepreneur Media’s filing called the construction what it is: “On that construction no one at Meta decided to torrent anything, yet Meta torrented shadow libraries repeatedly from 2022-2024.” [6]

    And the record contains Meta’s own contemporaneous assessment of the risk. An internal logbook, reviewed by an employee before a download from Anna’s Archive, reads:

    “Decided to go with direct file uploads without using torrents for the following reasons: Using torrents would entail seeding the file such as sharing content outside. This would be legally not okay.” [6]

    Hixson was unmoved by the consensus framing. “The rog isn’t limited to formal decisions,” he held — Meta’s answers had “wander[ed] away from the rog,” talking about use of the books when the question was who obtained or downloaded them. And the order turns the team-consensus defense inside out: “If a decision is fairly attributable to a group or a team, everyone in the group or team who participated in making that decision is responsive to this rog.” [7]

    His October 19 order requires Meta to identify everyone who took part in the discussion preceding the 2023 LibGen download, everyone who approved proceeding with BitTorrent, whoever directed or approved the April 2024 download from Anna’s Archive, and everyone who made, authorized or approved the decision “whether or how to seed, or to mitigate seeding” [6][7].

    The names matter because they bear on willfulness and on the contributory infringement and distribution claims [6] — and the distribution theory is the one that turned Bartz’s pirated downloads into a $1.5 billion settlement [2]. How much data Meta actually shared outward is still unclear: Meta’s own expert estimated that uploads from its servers during torrenting between April and July 2024 did not exceed roughly 30 percent of what was downloaded [6].


    The Evidence That Is Not There

    The missing command histories

    Back in March 2025, the court adopted a discovery definition that expressly included command-line history logs — the .bash_history files in the home directories of the servers used to download training data [5]. The rightsholders now say Meta produced none from the servers used during its torrenting periods until one surfaced in search-term results, that Judge Hixson ordered full compliance on September 11, 2026, and that they want the court to order Meta to state whether any such files were deleted, and when [5]. Meta’s position is that it has produced substantial logs from those servers, is still searching backups on a rolling basis, and that there is no basis to allege deletion [4].

    Why the shell history is decisive: those logs would show how Meta’s torrent clients were configured — including upload settings — and which torrent files the company targeted, which determines which works are in play [4]. A question of dataset provenance is being reconstructed from command lines because nothing else was written down at the time.

    The book club problem

    The rightsholders also want the contents of an internal Google Drive folder titled “ML Books” and records of a Meta book club that may have used books from shadow libraries. In one document Meta produced, an employee shared a “pirated book” with a colleague — “i know for vellum sake this is prob very ironic,” she wrote — and when asked whether the book was pirated, replied with a link to the ML Books folder. The recipient told another colleague, “bro she tryna get me fired,” adding that he “may need her to delete those messages” [4].

    The legal theory is per-use fair use. Under Andy Warhol Foundation v. Goldsmith, fair use is assessed separately for each use of a work — so books taken from a shadow library and then read for pleasure would constitute a second, unexcused use that sits outside any training-related defense [5]. The implication is quietly severe for every company that has assembled a corpus: the training defense covers training, not everything else employees do with the files afterward.

    Meta’s answer is a calendar. The two documents containing the links were produced August 17; the plaintiffs first asked for the linked files on October 9 — more than a month after fact discovery closed and past the local deadline for raising disputes [5]. On that reasoning the folder never gets produced, whatever is in it. The discovery warfare runs both directions: after the contributory claim was added, Meta served the plaintiffs 1,372 discovery requests on a single day — 84 to 94 per author — a demand the authors dismissed as a “complete do-over 18 months later” [3].


    The Paperwork Wall, and the Fee Clock

    While the acquisition fight draws the headlines, the quieter constraint on rightsholders is administrative. Entrepreneur Media’s amended complaint — due November 6, 2026 — was expected to add more than 400 additional books and more than 140 additional magazine issues, and almost none of them came from Entrepreneur’s own archive. They were identified in Meta’s interrogatory responses served between August 4 and September 30, and in the October 2 opening report of the rightsholders’ torrenting expert, which placed more than 400 Entrepreneur books inside the shadow-library collections Meta torrented [8].

    Only the defendant knows which works it acquired and trained on, so the operative claim emerges from the defendant’s own disclosures — twelve to thirty months after the original complaint, which is best understood as a placeholder [8]. And every added title must arrive with a U.S. copyright registration and a clean chain of title from author to claimant, because an infringement claim cannot be asserted on an unregistered work [8]. The back catalogue has to be registration-complete before the corpus list lands, not after.

    Which makes the timing of the Copyright Office’s first fee change since 2020 pointed. The proposed schedule, submitted to Congress on July 14, 2026, raises the basic registration fee from $65 to $85 — an increase the Office kept below its historical inflation benchmark to lift cost recovery from 47 percent toward 62 percent — with larger increases for renewals, group registrations, and appeals [9][10]. Under the statute, the Register may institute the new fees 120 days after submission unless Congress disapproves — a window that ran out in mid-November 2026, meaning the higher fees can take effect the same week the amended complaints are due [8][9]. For any publisher weighing a bulk catch-up registration program, the arithmetic changed this week.


    The Road to May 3, 2027

    Under the court’s October 14 order, Meta and Entrepreneur Media will prepare for trial on May 3, 2027, before Judge Chhabria, with the Kadrey authors’ case and three publisher actions — Cambronne, Chicken Soup for the Soul, and Cognella — queued behind it. If the lead case settles, the trial slot passes down in order of filing [5]. The queue is growing: Elsevier, Hachette, and Macmillan, along with more authors, joined the related-case list in October [4].

    The contrast with the parallel Anthropic track shows where the leverage now sits. The $1.5 billion Bartz settlement received final approval on July 20, 2026 — the largest copyright class settlement in history, covering roughly 482,000 eligible works at an estimated $3,000 per work before costs — even though training on lawfully acquired books was fair use in that case too [2]. The money was for the downloads. And the release is limited: works not on the court-approved Works List are not covered, so claims concerning them remain available [2].

    Chhabria, for his part, has already rejected the industry’s preferred framing of what a loss would mean. Calling the argument that liability would thwart a nascent technology “ridiculous,” he wrote: “these products are expected to generate billions, even trillions, of dollars for [developers]. If using copyrighted works to train the models is as necessary as the companies say, they will figure out a way to compensate copyright holders for it.” [1]

    Both tracks teach the same lesson from opposite directions. In Bartz, lawful training did not save Anthropic from paying $1.5 billion for unlawful acquisition. In Kadrey, a fair-use win on training has left Meta facing a 2027 trial over how — and through whom, and with what paper trail — roughly 196,000 pirated books reached its servers. The question that matters in AI copyright is no longer whether the models can learn from the corpus. It is how the corpus got there.


    What It Means

    • For AI developers: keep a provenance register, not a .bash_history. The alternative to recording the source, acquisition method, license or permission relied on, and approving manager for every corpus is reconstructing all of it from command lines under court order — with the gaps read against you [8].
    • Litigation holds must reach personal machines and former employees. The most damaging testimony in the October filing came from a researcher who had left the company three years earlier, describing a dataset copy deleted from his own computer [8][4].
    • Secondary internal uses can forfeit ground the training defense won. Because fair use is assessed per use, a book club that reads shadow-library copies is a separate, unexcused use — a reading group can lose what the training pipeline earned [5].
    • For rightsholders: build the claim from the defendant’s discovery, and register first. The original complaint is a placeholder; the real works list comes from the defendant’s own disclosures, and every added title needs a registration and clean chain of title in hand when the window opens — at a fee that just went up for the first time in six years [8][9].

    References

    [1] Skadden, Arps, Slate, Meagher & Flom LLP — “Fair Use and AI Training: Two Recent Decisions Highlight the Complexity of This Issue” (July 2025) [SECONDARY]

    [2] Heidrick, Jay, and Melenie Van (Polsinelli PC) — “AI vs. Authors Update: Court Approves Historic Anthropic Settlement While Meta Litigation Continues,” National Law Review (July 31, 2026) [SECONDARY]

    [3] Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal.) — Case docket including Plaintiffs’ Opposition to Meta’s Administrative Motion for Additional Discovery, Dkt. No. 797 (Aug. 4, 2026) [PRIMARY]

    [4] TorrentFreak — “Ex-Meta Researcher Torrented Pirated Books Dataset, Deleted It ‘Due to All This Litigation’” (October 2026) [SECONDARY]

    [5] Joint Case Management Statement, Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC, Dkt. No. 879 — U.S. District Court for the Northern District of California, filed October 27, 2026 [PRIMARY]

    [6] TorrentFreak — “Judge Orders Meta to Name Employees Who Approved Torrenting Anna’s Archive and LibGen” (October 2026) [SECONDARY]

    [7] Order re Interrogatory Responses, Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC — Magistrate Judge Thomas Hixson, U.S. District Court for the Northern District of California (October 2026) [PRIMARY]

    [8] PerspireIP — “Meta’s AI Training Copyright Trial Set for 3 May 2027 — and the Works List Isn’t Final” (November 1, 2026) [SECONDARY]

    [9] U.S. Copyright Office — “U.S. Copyright Office Submits Proposed Fee Schedule and Analysis to Congress,” NewsNet Issue 1090 (July 14, 2026) [PRIMARY]

    [10] U.S. Copyright Office — “Copyright Office Fees,” Federal Register (March 20, 2026) [PRIMARY]


    AI automated and generated content. Human reviewed.

  • The Consequence Test: How a First-Ever Examiner Framework, a Four-Tier Risk Ladder, and a September Renaming Brought “Super Intelligence” Into the Bank Exam

    The Consequence Test: How a First-Ever Examiner Framework, a Four-Tier Risk Ladder, and a September Renaming Brought “Super Intelligence” Into the Bank Exam

    On November 4 the OCC issued the first federal framework telling bank examiners how to probe AI use — and its central move is to refuse what everyone expected: no separate AI risk category, no separate risk appetite, just consequences, autonomy, and oversight.


    What the Framework Is — and Isn’t

    OCC Bulletin 2026-54, “Super Intelligence: Supervisory Framework,” was issued November 4, 2026 to the chief executives of all national banks, federal savings associations, and federal branches, along with every OCC examining employee [1]. It is a supervisory framework designed to support OCC examiners in evaluating banks’ use of super intelligence (SI) — the term the executive branch now uses in place of “artificial intelligence” — in normal-course business activities [1]. The OCC described it as the first-ever guidance for how examiners should probe banks’ AI usage [2].

    The document’s own framing matters as much as its content. It does not establish supervisory expectations for banks, sets no enforceable standards or prescriptive requirements, and banks are not expected to use it to assess their own SI risk management [1]. Examiners are told to tailor their work to the materiality of each SI use case, proportionate to the bank’s size, complexity, and risk profile — language the bulletin repeats in a dedicated note for community banks [1]. Reuters notes the practical caveat: the document is not legally binding like a rule, but banks typically comply closely with such guidance [2]. An examiner resource that disclaims enforceability still becomes the script for the exam.

    The scope of activities covered is the full stack of modern banking: credit decisions, payments, fraud detection, customer service, operations, and risk management, with the OCC intending to update the framework as the technology and supervisory understanding mature [1][3]. Banking regulators worldwide have been querying lenders about AI use over cybersecurity and financial-stability concerns, and US regulators have spent 2026 asking banks what advanced models they run, what limits they place on those systems, and how they monitor them [2].


    The Consequence Test

    The framework’s core move is a rejection of technology-specific risk. Its highlights state plainly:

    “The issuance clarifies that examiners should not treat SI as a separate and distinct risk category, and banks are not required to define a separate SI risk appetite.” [1]

    That single sentence reverses the default assumption in most AI-governance regimes — that AI is a novel risk deserving its own taxonomy, its own committee, its own appetite statement. The OCC’s logic is the opposite: SI introduces, amplifies, or modifies risk inside categories banks already own — credit, operational, compliance, model risk — so supervision should follow the existing risk, not invent a new silo [4]. The framework groups 16 components into six pillars — purpose, risk, control, oversight, resilience, and assurance — and adds an illustrative control-risk tiering methodology plus a technical dictionary [4].

    The second move follows from the first. Risk classification, the bulletin says, may be based on potential consequences, transfer of agency, and the degree of human oversight — “rather than the technical sophistication of the SI used” [1]. Control intensity could increase with the potential consequences of an agent’s failure or misuse [1]. The OCC’s example makes the inversion concrete: a simple AI agent authorized to initiate transactions could be considered riskier than a sophisticated model deployed for internal purposes, and higher-risk uses may require greater restrictions on autonomy or more direct human oversight [2]. The question an examiner is trained to ask is not “how advanced is your model?” but “what can it do, and who can stop it?”

    Analysts read the consequence test as the framework’s center of gravity. Because examiners scope by materiality, attention will concentrate on the highest-consequence, most autonomous use cases — and governance focused only on advanced models may leave risk misaligned [4]. The absence of a separate risk category, in other words, does not mean lighter scrutiny; it means scrutiny aimed at where AI touches money and customers rather than at the technology itself.


    A Four-Tier Ladder From Payment Agents to Office Chatbots

    The illustrative tiering methodology is the framework’s most operational feature — the closest thing to a checklist the document contains. The tiers run [4]:

    • Tier 1, Critical. Direct action, low or no reversibility, high financial or customer impact, autonomous or partially autonomous. Illustrative cases: autonomous payment agents, real-time credit decisioning, fraud-hold release, and market access. The common thread is that the system moves money or gates access without a human in the loop — the use cases where an error compounds faster than a person can react.
    • Tier 2, High. Direct action or high-weight decision support, human-on-the-loop, moderate-to-high impact. Illustrative cases: SI-assisted underwriting, agent-initiated customer communications, AML alert disposition, and code assistance for internal code or data access changes. Here a human still signs off, but the machine has already shaped the decision space — the risk shifts from the action to the quality of the oversight.
    • Tier 3, Moderate. Decision support with mandatory human-in-the-loop, limited data sensitivity, reversible. Illustrative cases: internal document summarization for credit memos and agent-assisted research. The machine drafts; a person decides; the output can be redone.
    • Tier 4, Low. No customer or financial data exposure, no direct action, fully human-controlled use. Illustrative cases: internal productivity tools, code assistance, and general employee Q&A.

    Read as a ladder, the tiers encode a supervisory philosophy: autonomy and irreversibility are the risk multipliers, not intelligence. A bank could run a frontier-grade model at Tier 4 and a rules-based script at Tier 1, and the framework tells examiners the script deserves the harder look [2][4]. That inverts the incentive structure most governance programs have built — which tend to inventory and review models by technical capability rather than by what the system is authorized to do.

    Consultants reviewing the framework flag where preparation effort should concentrate: inventory and purpose (activity outside an approved purpose may be treated as a control failure even when the output looks reasonable), human oversight that is more than a click (bulk approvals and alert fatigue erode its value), resilience (kill switches, tested fallbacks, and a customer remediation process — because an agent can stay available while taking harmful actions), third parties (accountability remains with the bank), and independent assurance that can reconstruct a material decision [4]. Each of those maps back to a pillar rather than a technology.


    The Gap It Fills

    The framework lands in a hole the OCC and its fellow regulators dug themselves. On April 17, 2026, the OCC, Federal Reserve, and FDIC issued revised interagency model-risk guidance — OCC Bulletin 2026-13, the successor to the 2011 guidance that had served as the banking system’s de facto AI rulebook — and stated plainly that “generative AI and agentic AI models … are not within the scope of this guidance,” while promising a future request for information on precisely those tools [5]. As of November, that RFI remains unpublished. The April revision is also principles-based, written to remove prescriptive control expectations and most relevant to banks above $30 billion in assets [5].

    The June 25 credit-examination update tightened the same knot from the other side. OCC Bulletin 2026-29 revised the “Lending and Loan Portfolio Risk Management” booklet of the Comptroller’s Handbook, embedding model risk management, third-party risk management, and automated credit approval processes directly in credit examinations — yet the words “artificial intelligence” and “machine learning” do not appear in the text [5]. The result was a supervisory white space: examiners were being pointed at AI-driven lending through guidance that did not name the technology, while the model-risk guidance that would logically cover it excluded the newest tools [5]. Bulletin 2026-54 is the OCC’s answer — an examiner-facing resource for exactly the generative and agentic systems the model-risk guidance leaves out, though the agencies still plan to seek input that could shift that boundary [4][5].

    Two limits are worth stating plainly. The framework covers OCC-supervised institutions only, so state-chartered banks should watch the Federal Reserve and FDIC positions [4]. And because it is not a self-assessment tool, nothing in it legally requires a bank to do anything — but every examination conversation from here forward will be shaped by a document the bank never had to write [4].


    The Name on the Door

    The bulletin’s vocabulary is itself a policy artifact. “Super Intelligence” refers to “the technologies and systems previously encompassed by the term ‘Artificial Intelligence’” as defined in 15 USC 9401(3) — a definition imported wholesale from Executive Order 14434, “Inaugurating the Era of Super Intelligence,” signed September 29, 2026 and published October 2 [1][6]. The order directs executive agencies to use “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI” in official correspondence, public communications, websites, and policy documents, to the maximum extent permitted by law [6]. Its stated reason:

    “The capabilities of today’s frontier systems do much more than imitate or automate discrete aspects of human intelligence. They increasingly amplify human ingenuity and unlock new forms of creativity … Accordingly, the term ‘Super Intelligence’ more appropriately captures the promise, potential, and rapidly advancing capabilities of these technologies.” [6]

    The redefinition is not merely cosmetic. Within 60 days of the order, the Assistant to the President for Science and Technology must submit proposed legislative language to establish a federal definition of SI — language that could modify, expand upon, or supersede the existing statutory definition of artificial intelligence, with conforming amendments to every statute that uses the old term [6]. Only Congress can change the statutory definition, and the proposal is due in late November [6]. In the meantime, every federal document — including a bank examiner’s framework — speaks a renamed language for the same technology. Whether “super intelligence” eventually means something legally different from “artificial intelligence” is one of the open questions this order set in motion.


    What to Watch

    • The exam cycle. The framework is live for examiners now, and Reuters reports regulators have already been asking banks about advanced-model limits and monitoring [2]. The first full examination cycle run under the tiering logic — especially how examiners treat autonomous payment agents — will show whether consequence-based scoping concentrates or diffuses scrutiny in practice.
    • The promised RFI. The April model-risk guidance excluded generative and agentic AI pending a request for information that still has not appeared [4][5]. If the interagency RFI lands, the boundary between model-risk guidance and this framework could be redrawn.
    • The Fed and FDIC. The framework binds OCC examiners only; state-chartered banks await parallel positions from the Federal Reserve and FDIC [4]. Whether the other agencies adopt the same consequence tiers or write their own will determine whether this becomes the industry standard.
    • The statutory redefinition. The EO 14434 legislative proposal on a federal SI definition is due in late November [6]. If it modifies 15 USC 9401(3), every regulation built on the old definition — including this framework’s own footnote — inherits a new substrate.
    • Updates to the framework. The OCC says it will update the examiner guide as technology, use cases, and supervisory understanding change [1]. For a document that disclaims enforceability, version history will do a lot of the policy work.

    The deeper design choice is the one worth naming. Most AI-governance regimes build a special box for the technology and then argue about what goes in it. The OCC did the reverse: it dissolved the box and told examiners to grade what the technology can do to the bank — move money, gate credit, release fraud holds — under risk categories that existed before the technology did. If that approach survives contact with the first few exam cycles, it will be because the consequence test answers the question that matters at a bank: not how smart the system is, but who pays when it is wrong.


    References

    [1] Office of the Comptroller of the Currency — [PRIMARY] “Super Intelligence: Supervisory Framework.” OCC Bulletin 2026-54. November 4, 2026.

    [2] Reuters — [SECONDARY] Schroeder, Pete. “US bank regulator issues first-ever AI guidance for examiners.” November 4, 2026.

    [3] ABA Banking Journal — [SECONDARY] “OCC releases framework for supervising artificial intelligence.” November 4, 2026. (American Bankers Association publication)

    [4] Forvis Mazars — [SECONDARY] “OCC AI Framework: Consequences, Not Sophistication, Drive Scrutiny.” November 2026. (accounting and advisory firm analysis)

    [5] FTI Consulting — [SECONDARY] Amerasinghe, Christopher, and Kyle Wetzold. “AI Risk Considerations Move into OCC Credit Examinations.” October 27, 2026. (consulting firm analysis)

    [6] The White House — [PRIMARY] “Inaugurating the Era of Super Intelligence.” Executive Order 14434, 91 Fed. Reg. 63129. September 29, 2026.


    AI automated and generated content. Human reviewed.


    Listen to the Podcast: AI Policy for the Bears

    Episode 44: The Consequence Test: How a First-Ever Examiner Framework, a Four-Tier Risk Ladder, and a September Renaming Brought “Super Intelligence” Into the Bank Exam — click to listen

    Two bears break down this story on AI Policy for the Bears, Episode 44.

  • The Vanishing Model Card: How a December 28 Deadline, a $1.55-Billion Estimate, and a Majority Opposition Ended the Federal Transparency Floor for Clinical AI

    The Vanishing Model Card: How a December 28 Deadline, a $1.55-Billion Estimate, and a Majority Opposition Ended the Federal Transparency Floor for Clinical AI

    HHS’s HTI-5 final rule, published October 28, retires the algorithm-transparency requirements its own HTI-1 rule created two years ago — over the written objections of most commenters — at the same moment state legislatures are writing human-in-the-loop requirements for health AI into law.


    What the Rule Does

    The Health Data, Technology, and Interoperability rule known as HTI-5 — formally “ONC Deregulatory Actions To Unleash Prosperity” — was published in the Federal Register on October 28, 2026 and takes effect December 28, 2026, with three amendatory instructions deferred to January 1, 2027 [1]. It is an explicit implementation of Executive Order 14192, “Unleashing Prosperity Through Deregulation” [1]. HHS projects the changes will save $1.55 billion in present-value regulatory costs, with $108.1 million in annualized net savings beginning in 2027 [2]. Of the 60 certification criteria in the ONC Health IT Certification Program, the rule removes 14 and revises seven, leaving 39 unchanged; the original proposal had targeted 34 removals, and the final version scaled the overhaul back substantially [2].

    “American doctors and patients need better technology, not more Washington paperwork.” — HHS Secretary Robert F. Kennedy Jr. [2]

    The most consequential line item for AI policy sits inside the decision support interventions (DSI) certification criterion at 45 CFR 170.315(b)(11). The rule finalizes removal of paragraph (iv) — the “source attribute” transparency data, the model-card requirements — along with paragraph (v), which governed access to and modification of that information, and paragraph (vi), the intervention risk management requirement for predictive DSIs [1]. Because those requirements are gone, the rule also strikes the parallel maintenance obligation at 170.402(b)(4), which required developers to keep source attribute information current [1]. National Coordinator for Health IT Tom Keane framed the practical effect as freeing developers: time and money once spent complying with requirements “that are no longer necessary” can go toward interoperability and automation work instead [2].


    The Transparency Framework Being Retired

    The requirements being removed are barely two compliance-years old. The HTI-1 final rule, published January 2024, was described at adoption as the first federal regulatory approach establishing transparency requirements for AI and other predictive algorithms built into certified health IT [3]. Its core mechanism was the “source attribute”: plain-language disclosures attached to decision support tools — 13 attributes for evidence-based interventions and 31 for predictive ones — covering development details, input features, demographic representativeness, fairness and validity metrics, and maintenance schedules [3]. Developers of certified health IT have had to comply with the associated maintenance requirement since January 1, 2025 [3].

    The underlying logic was procurement-level scrutiny. A clinician or a hospital buying an EHR with an embedded prediction tool could look up how the model was trained, on whom, how it performed on local data, and how often it was re-validated — the raw material for judging whether a tool was FAVES: fair, appropriate, valid, effective, and safe [3]. It was a disclosure regime, not a performance regime. Nothing in HTI-1 required a model to be good; it required only that its makers document it in a standard, inspectable way. The bet was that sunlight at the point of purchase would discipline the market without a regulator ever grading the models.


    The Comment Record and the Utility Argument

    The rule’s own preamble is unusually candid about the opposition. “A large majority of commenters were not supportive” of removing the DSI transparency and risk-management requirements, citing concerns that the change would shift evaluation, governance, and liability burdens to providers and health systems and undermine patient safety, clinician trust, and the ability to evaluate bias, performance, and appropriateness [1]. Over half of the commenters with transparency concerns asked ONC to retain the model-card requirements or the entire criterion [1].

    ONC’s answer is a utility argument, not a safety argument. The agency wrote that developers reported most clinical users “lack the time or technical background to engage with source attribute information” at the point of care, and that ONC identified no publicly available evidence the requirements led to positive impacts on patient care, including the removal of defective algorithms [1]. In the agency’s conclusion, “the requirements lacked utility” [1]. Two things follow from that reasoning. First, the burden of evaluating clinical AI shifts from a federal disclosure floor to whoever buys it — hospitals, health systems, and clinicians negotiating alone. Second, ONC left a door open: it committed to “further consider for future rulemaking whether these or similar requirements should be reinstated” [1].

    What survives in the criterion matters too. Requirements supporting DSI selection and the electronic feedback tools for evidence-based interventions were retained — ONC notes these let providers share, export, and report on real-world DSI performance [1]. A hospital can still measure how a tool behaves in its own data; what disappears is the standardized dossier describing how the tool was built.


    The States Are Moving the Other Way

    While HHS deregulates the tools, state legislatures are regulating the uses. So far in 2026, states have passed 14 new laws regulating AI in healthcare across 11 states — seven states limited AI in medical authorization decisions and five banned AI-delivered therapy services [4]. The insurer laws share one principle: AI may assist the workflow, but a licensed human must own the adverse decision.

    • Alabama SB 63 (effective October 1, 2026) mirrors CMS’s Medicare Advantage guardrails: AI-based prior authorization determinations must rest on the patient’s medical history and unique clinical circumstances as presented by the treating provider, and insurers must certify annually that their AI does not rely on group datasets, does not discriminate against subscriber groups, and is monitored for accuracy [5].
    • Indiana HB 1271 (effective July 1, 2026) reaches beyond prior authorization: insurers cannot use AI as the sole basis for downcoding a claim without a professional reviewing the medical record, and providers cannot submit AI-generated claims without review [5].
    • Utah SB 319 (effective January 1, 2027) requires insurers to disclose AI use in prior authorization review to the state, providers, and enrollees, and requires anyone reviewing an adverse determination to exercise independent medical judgment rather than relying on the tool’s recommendation [5].
    • Washington SB 5395 (enacted March 2026) bars AI as the sole means used to deny, delay, or modify care, reserves adverse determinations for licensed physicians or health professionals, and requires carriers to report to the insurance commissioner how many denials involved AI [5].
    • Maryland HB 1563 (effective June 1, 2026) adds quarterly reporting of adverse decisions and whether AI was involved, and empowers the insurance commissioner to investigate spikes — with particular attention to emergency department denials [5].
    • Georgia SB 544 (effective January 1, 2027) expressly permits AI in the prior authorization process to automate tasks and reduce administrative burden, but prohibits adverse determinations issued without the review and approval of a licensed healthcare provider [5].

    The pattern generalizes beyond insurers. Five states enacted therapy chatbot bans in 2026 — Colorado, Maine, Tennessee, Vermont, and Rhode Island — building on the Illinois and Nevada bans of 2025 [4]. Delaware HB 191 bars any nonhuman entity, including an AI agent, from being licensed or certified as a nurse, physician, or physician assistant, or from using protected professional titles [5]. Arizona’s behavioral health board will require documented informed consent for AI-involved services starting January 1, 2027 [5].

    The two tracks do not directly collide, because HTI-5 governs what vendors must tell their customers while the state laws govern what insurers and clinicians may do with the technology. But the direction of travel is unmistakable: less mandated visibility into the model, more mandated accountability for the decision. Notably, the state laws mostly regulate the payer side — the utilization-review algorithms whose denials generated the litigation and congressional heat — while the federal disclosure floor being retired covered the provider side, where the harms are quieter and the buyers are large institutions with negotiating power.


    Why the Stakes Go Beyond Paperwork

    The transparency floor being retired sits on the provider side of a system whose payer side is already under litigation and congressional scrutiny. More than half of eligible Medicare beneficiaries — 54 percent — are enrolled in Medicare Advantage plans administered by private insurers [6]. A February 2026 peer-reviewed analysis in npj Digital Medicine documents the pattern: a congressional report concluded that denial rates in UnitedHealthcare, CVS, and Humana MA plans jumped significantly after increased use of AI [6]. The same analysis describes the Lokken class action over UnitedHealth’s nH Predict algorithm — which plaintiffs allege carried a 90 percent error rate from lack of human review, with over 80 percent of its prior authorization denials reversed on appeal — and notes that the federal judge allowing part of the case to proceed signaled that the measuring stick will be the insurer’s own evidence-of-coverage promise that decisions would be made by “clinical services staff” and “physicians” [6].

    Washington has noticed across party lines. In July 2026, Sens. Richard Blumenthal (D-Conn.) and Josh Hawley (R-Mo.) wrote to the executives of UnitedHealth Group, Humana, and CVS Health demanding internal records on their use of AI to block rehabilitative care, one month after an HHS Office of Inspector General report found a continuing pattern of denials that, in the senators’ view, undercut the insurers’ claims to have reduced barriers to care [7]. That is the backdrop against which the model-card requirement disappears: the same federal department publishing the deregulatory rule is, through its inspector general, documenting the denial patterns that motivated transparency in the first place.


    What to Watch

    • December 28 versus January 1. The federal transparency requirements lapse December 28, 2026, while several of the state insurer laws (Utah, Georgia) and Arizona’s consent rule take effect January 1, 2027 [1][5]. For the first time in this space, the strictest written constraints on health AI will be state law rather than federal certification.
    • HTI-6 and information blocking. The December 2025 proposals to redefine “access” and “use” so autonomous AI systems can retrieve and share health data were left undecided and “remain under consideration”; a draft HTI-6 rule covering APIs and information blocking updates sits on the 2026 Unified Agenda [2]. Whether AI agents can lawfully pull patient records may be decided there.
    • The promised reconsideration. ONC committed to consider reinstating source-attribute transparency in future rulemaking [1]. Whether that happens, and in what form, will show whether HTI-5 is a revision of HTI-1’s approach or a repudiation of it.
    • The terminology shift. Under Executive Order 14434, federal agencies are replacing “artificial intelligence” with “super intelligence” in official documents, and the HTI-5 preamble already speaks of “SI-enabled” decision support [1][6]. The vocabulary of health AI policy is being rewritten alongside its rules.

    The deeper trade is the one the comment record named. HTI-1’s premise was that a clinician cannot safely delegate judgment to a model nobody can inspect. HTI-5’s premise is that nobody was inspecting it anyway — so the inspection obligation, not the model, is what gets removed. What replaces the federal floor is a patchwork of state licensure walls, insurer-disclosure statutes, and whatever procurement diligence hospitals can negotiate on their own. Whether that produces better tools, as HHS promises, or simply quieter ones, is the question the next year of denials and appeals will answer.


    References

    [1] Federal Register — [PRIMARY] “Health Data, Technology, and Interoperability: ONC Deregulatory Actions To Unleash Prosperity.” 91 FR 67916. U.S. Department of Health and Human Services. Final rule, published October 28, 2026.

    [2] Fierce Healthcare — [SECONDARY] Landi, Heather. “HHS finalizes rule to ease health IT certification requirements but punts info blocking proposals.” October 26, 2026.

    [3] ONC/ASTP, U.S. Department of Health and Human Services — [PRIMARY] “HTI-1 Decision Support Interventions (DSI) Fact Sheet.” January 2024.

    [4] Transparency Coalition — [ADVOCACY] Barcott, Bruce. “States have passed 14 new laws this year regulating the use of AI in healthcare.” July 27, 2026. (advocacy organization)

    [5] Holland & Knight — [SECONDARY] Silverboard, Dan M., Jacob Anthony Barrera, and Isabella Ryan. “States Continue Efforts to Regulate AI in Healthcare: A Review of Legislation Passed in 2026.” May 26, 2026. (law firm alert)

    [6] npj Digital Medicine (via PubMed Central) — [SECONDARY] Raza, Sara, Sara Gerke, Christina Silcox, Rachele Hendricks-Sturrup, and Carmel Shachar. “Medicare advantage becoming a disadvantage with use of artificial intelligence in prior authorization review.” npj Digital Medicine 9:208. February 4, 2026.

    [7] STAT News — [SECONDARY] Ross, Casey, and Bob Herman. “Medicare Advantage insurers face new bipartisan scrutiny over AI and care denials.” July 15, 2026.


    AI automated and generated content. Human reviewed.


    Listen to the Podcast: AI Policy for the Bears

    Episode 43: The Vanishing Model Card: How a December 28 Deadline, a $1.55-Billion Estimate, and a Majority Opposition Ended the Federal Transparency Floor for Clinical AI — click to listen

    Two bears break down this story on AI Policy for the Bears, Episode 43.

  • The Borrowed Notice: How a Six-Week Repeal, a Vanished Bank Exemption, and a Rulemaking Hearing on Its Final Comment Day Made Colorado the First State AI Law to Bind Banks Directly

    The Borrowed Notice: How a Six-Week Repeal, a Vanished Bank Exemption, and a Rulemaking Hearing on Its Final Comment Day Made Colorado the First State AI Law to Bind Banks Directly

    Colorado repealed its first-in-the-nation AI law on May 14, 2026 — six weeks before it would have taken effect — and replaced it with a disclosure regime that quietly drops the exemption that kept banks and credit unions outside its scope. The centerpiece trade: a lender’s existing federal adverse-action notice now does double duty as state compliance, while the anti-bias obligations that once sat on both sides of that trade are gone.


    A Six-Week Repeal

    In 2024, Colorado enacted Senate Bill 24-205, the first comprehensive state AI law in the United States, built around “high-risk artificial intelligence systems” and a duty to prevent algorithmic discrimination.[1] It was set to take effect June 30, 2026. On May 14, 2026 — with days left in the legislative session — Governor Jared Polis signed Senate Bill 26-189, repealing and replacing the statute “mere weeks before the original statute’s June 30, 2026 effective date,” in the words of one legal analysis.[2] The replacement law, the Automated Decision-Making Technology (ADMT) Act, takes effect January 1, 2027.[2]

    The rewrite changes the law’s basic vocabulary. Gone are “high-risk AI systems”; in their place is “automated decision-making technology” — any technology that processes personal data and uses computation to generate outputs used to make, guide, or assist a decision about an individual.[1] The law applies only when ADMT “materially influences” a “consequential decision,” meaning the system’s output is a non-de minimis factor that affects the outcome, including by ranking, scoring, or recommending.[2] Consequential decisions include access to or eligibility for education, employment, housing, healthcare, insurance, essential government services — and financial or lending services.[1]

    The design logic behind the swap is a bet on regulation by disclosure rather than regulation by governance. Instead of requiring firms to test systems for bias before deployment and maintain formal risk programs, the state requires them to tell consumers when ADMT shaped a decision and give them a route to challenge it after the fact.[2] Whether post-decision transparency can substitute for upstream testing — or merely relocates the burden from institutions to consumers — is the question the rest of the law turns on.


    What Died, What Survived

    The replacement eliminates most of the 2024 law’s substantive obligations:[2]

    • The anti-discrimination duty. Developers and deployers no longer must use “reasonable care” to protect consumers from algorithmic discrimination — the concept the 2024 law was built around. The replacement statute “does not mention algorithmic discrimination.”[2]
    • Impact assessments and risk management programs. Deployers no longer complete annual impact assessments or maintain formal risk-management policies for covered systems.[2]
    • Attorney general reporting. The 90-day duty to notify the attorney general after discovering algorithmic discrimination is gone.[2]
    • The bank and credit union exemption. The provision letting regulated financial institutions claim full compliance through prudential oversight is eliminated entirely.[3]

    What survives is a procedural shell: deployers must give clear, conspicuous notice at the point of interaction; provide a plain-language description of the decision and the ADMT’s role within 30 days of an adverse outcome; honor consumer rights to access and correct inaccurate personal data; offer meaningful human review and reconsideration “to the extent commercially reasonable”; and retain compliance records for three years.[1][2] Enforcement belongs exclusively to the attorney general under the Colorado Consumer Protection Act, where a violation is a deceptive trade practice, with a 60-day cure opportunity before January 1, 2030 when a cure is possible — and no new private right of action.[1]


    The Exemption That Vanished

    For banks and credit unions, the change cuts in a direction most coverage of the repeal missed. The 2024 law contained an exemption under which banks, credit unions, and their affiliates could claim full compliance if they were subject to a state or federal prudential regulator applying requirements substantially equivalent to the statute’s.[3] SB 189 “entirely eliminates that exemption.”[3]

    Nor does the replacement contain an entity- or data-level exemption tied to the Gramm-Leach-Bliley Act, the federal privacy statute that governs financial institutions and frequently shields them from state privacy laws.[3] The practical consequence: a bank doing business with Colorado consumers that uses scoring or ranking technology in a credit decision is now directly regulated by state AI law for the first time. As one law firm analysis put it, the unambiguous pull of regulated entities into scope “marks a new era in state AI laws for financial institutions.”[3]

    The law does carve out functions regulators did not want to burden: technologies used for anti-money-laundering and counter-terrorist-financing controls, sanctions compliance, and fraud prevention, including identity verification.[3] That split reflects a judgment about where automation is protective rather than consequential — screening for fraud and money laundering affects whether a transaction is permitted, not whether a consumer gets credit on fair terms. The line the statute draws runs between the security perimeter and the credit decision itself.


    The Borrowed Notice

    The statute’s key accommodation for lenders is a safe harbor that converts their existing compliance machinery into state compliance:

    “A financial institution that is required to provide, and does provide, a notice to a consumer under the Equal Credit Opportunity Act (ECOA) and its implementing Regulation B, and where applicable under the Fair Credit Reporting Act (FCRA), complies with the notice and disclosure requirements of SB 189 for the same decision or adverse outcome.”[3]

    The core logic is de-duplication. Federal law has required creditors to send adverse-action notices stating specific reasons for denials since 1974 — an obligation that, as the Center for Democracy and Technology (a digital-rights advocacy organization) summarizes the statutory baseline, applies “even when AI was used in any aspect of the credit decision, regardless of the AI’s complexity or opacity.”[4] Rather than layering a second, state-specific notice on top, Colorado treats the federal notice as sufficient for the same decision.[3] For banks, the incremental compliance work therefore concentrates in what federal law never required: the data-correction and human-review workflows triggered after an adverse outcome.

    The timing makes the trade stark. On April 22, 2026, the CFPB finalized a rule amending Regulation B to remove disparate-impact liability from federal credit law.[4] With effects-based enforcement gone at the federal level and the reasonable-care anti-bias duty gone from Colorado’s statute, what remains protecting a consumer denied credit by an algorithm in Colorado is the notice explaining why, the right to correct bad inputs, and the right to a human second look.[3][4] The state’s bet is that disclosure plus human review can do the work that upstream bias testing used to do — a proposition no one has yet tested.


    Today’s Hearing, and What the Rules Would Require

    The attorney general’s rulemaking has moved in parallel. Proposed rules (4 C.C.R. 904-6) were filed with the Secretary of State on August 11, 2026; an interim draft followed on October 6; and a regulatory analysis was released October 21.[5] The formal rulemaking hearing — in person at 1300 Broadway in Denver and virtually — is scheduled for October 26, 2026, the same day the written comment period closes at 11:59 p.m. Mountain Time.[5]

    The draft rules would translate the statute’s general duties into operational deadlines:[6]

    • Notice design. Post-adverse-outcome notices would go out through two channels where feasible, including an online method, and must explain the decision, the ADMT’s role, the human involvement, the principal reasons, relevant scores or inferences, and how to exercise rights.[6]
    • Response deadlines. Instructions for submitting rights requests within 24 hours of a request; responses to data and correction requests within 45 days; acknowledgment of human-review requests within 10 days; completion of human review within 45 days.[6]
    • Supply-chain documentation. A new “midstream developer” role would have to obtain upstream documentation and pass it through the AI supply chain to deployers.[6]

    The most consequential drafting choice is what the rules say a “human” review must be. The reviewer cannot be the original decision-maker, should sit outside the decision-maker’s reporting chain where feasible, must have subject-matter knowledge and the authority to approve, modify, or override the decision, and must be protected from retaliation — and, critically, “ADMT may not assist in the human review itself.”[6] That last clause is the difference between a genuine second look and a rubber stamp run past the same model. The rules would also presume review to be commercially reasonable when the adverse outcome is a severe and irreversible denial of a basic human need, subject to a demanding rebuttal.[6]


    The Shadow Over January 1

    Whether the law takes effect on schedule depends on litigation that predates the replacement. On April 9, 2026, xAI — now part of SpaceX — sued in the U.S. District Court for the District of Colorado to enjoin the 2024 law, raising First Amendment, Commerce Clause, vagueness, and equal-protection claims; the Department of Justice intervened on April 24, citing an executive order that singled out the Colorado law as problematic state AI regulation.[6] On April 27, the court entered a stipulated order under which the attorney general agreed not to enforce or investigate the 2024 statute while the case pends.[6]

    The same order preserves the challenge’s reach into the new law: xAI may move for a preliminary injunction within 28 days after final adoption of rules implementing the replacement.[6] That means the hearing that closes today starts a 28-day clock toward a possible federal injunction against a law that has not yet taken effect — an unusual posture in which compliance planning and constitutional litigation run on the same calendar, and lenders cannot prudently wait for either to resolve.[6]


    What to Watch

    • The “materially influence” gap. The draft rules declined to supply the presumptions and safe harbors stakeholders requested, leaving common screening and ranking tools to fact-specific analysis — the question that determines how many systems fall in scope at all.[6]
    • Whether xAI moves to enjoin. The 28-day window after final rule adoption is the last clear checkpoint before the January 1, 2027 effective date.[6]
    • The first enforcement signals. The attorney general must report annually on enforcement, and the 60-day cure period gives banks a preview of what the office considers a violation before penalties attach.[1]
    • The ECOA safe harbor in practice. Whether Colorado treats an adequate federal adverse-action notice as fully dispositive — or as a floor that the human-review deadlines sit on top of — will decide how much net-new work the law creates for lenders.[3]

    References

    [1] [PRIMARY] Colorado General Assembly. “SB26-189: Automated Decision-Making Technology.” Signed May 14, 2026. leg.colorado.gov

    [2] [SECONDARY] Schwartz, Tessa, Marian A. Waldmann Agarwal, and Joshua R. Fattal. “Colorado Hits Reset on AI Regulation With a New AI Act: What Developers and Deployers Need to Know.” Morrison Foerster. May 15, 2026. mofo.com

    [3] [SECONDARY] “The New Colorado AI Act: What Financial Institutions Need to Know.” Cooley Finsights. May 2026. finsights.cooley.com

    [4] [ADVOCACY] Center for Democracy and Technology. “AI in Financial Services: Current Legal Landscape.” 2026. cdt.org

    [5] [PRIMARY] Colorado Attorney General. “Colorado Automated Decision-Making Technology & Chatbot Safety Rulemaking.” Accessed October 26, 2026. coag.gov

    [6] [SECONDARY] McGinnis, Karin M., and Tiffany E. Payne. “Colorado’s Proposed ADMT Rules: What Businesses Need to Know Now.” Moore & Van Allen. October 5, 2026. mvalaw.com


    AI automated and generated content. Human reviewed.


    Listen to the Podcast: AI Policy for the Bears

    Episode 39: The Borrowed Notice: How a Six-Week Repeal, a Vanished Bank Exemption, and a Rulemaking Hearing on Its Final Comment Day Made Colorado the First State AI Law to Bind Banks Directly — click to listen

    Two bears break down this story on AI Policy for the Bears, Episode 39.

  • The Gray Zone: How a $10,000 Penalty, a Shut-Down Trial, and Five Copycat Bans Split State AI Therapy Law From the Science

    The Gray Zone: How a $10,000 Penalty, a Shut-Down Trial, and Five Copycat Bans Split State AI Therapy Law From the Science

    Illinois wrote the nation’s first statute defining AI’s legal role in psychotherapy in August 2025, and five states copied the approach in 2026. But the same bans that block predatory companion apps have no pathway for clinically validated tools — and the researchers who ran the first randomized trial of an AI therapy chatbot shut down their next study in Illinois rather than risk the law.


    Three Buckets and a Hard Line

    In August 2025, Illinois enacted Public Act 104-0054 — the Wellness and Oversight for Psychological Resources Act — described in a peer-reviewed legal analysis as the first state statute in the United States to explicitly define and regulate the use of AI in psychotherapy services.[1][2] The law sorts every AI function a therapist might reach for into three categories. Administrative support covers scheduling, billing, and general communications. Supplementary support covers record-keeping and progress tracking, with session recordings or transcripts requiring the patient’s explicit written consent. The third category, therapeutic communication, is defined as any interaction intended to diagnose, treat, or address a person’s mental, emotional, or behavioral health concerns.[1]

    Then the law draws the line. A licensed professional may not allow AI to make independent therapeutic decisions, interact with clients directly, generate treatment recommendations for the professional to sign, or detect emotions or mental states — and these prohibitions hold even when the patient consents. Violations are enforced by the Illinois Department of Financial and Professional Regulation at up to $10,000 per violation.[1][2] The core logic is licensure: therapy may be delivered only by someone answerable to a licensing board, and no chatbot holds a license, answers to a board, or carries malpractice accountability.

    In 2026, five more states enacted their own AI therapy restrictions: Colorado (enacted June 3, effective August 12, 2026), Maine (enacted April 13, effective July 29, 2026, where offering AI therapy without a licensed professional counts as an unfair trade practice), Rhode Island (enacted June 22, effective January 1, 2027), Tennessee (enacted April 1, effective July 1, 2026, barring AI systems that advertise themselves as qualified mental health professionals), and Vermont (effective June 17, 2026) — joining Illinois and Nevada, which adopted similar measures the year before.[3]


    The Gray Areas Nobody Drew Lines Around

    A legal analysis published in JMIR Mental Health by researchers at Rush University Medical Center mapped what the Illinois statute permits, prohibits, and leaves undefined — and the undefined middle is where the practical trouble sits.[2] The gray zone includes four recurring cases:

    • Risk-flagging tools that alert a therapist when a patient’s language in session may signal risk of self-harm
    • Session summaries and psychoeducation drafted by AI that could be read as therapeutic advice
    • Between-session check-ins that text patients about mood or therapy homework
    • AI-aided intake tools that straddle administrative and diagnostic functions

    Each of these could plausibly count as supplementary support — permissible — or as therapeutic communication or emotion detection, both banned outright. The authors’ sharpest example is the safety case. Research groups have proposed AI that listens to live or recorded sessions and flags clinicians to conduct additional suicide-risk assessment based on a patient’s voice or word choice. Does that mean the AI detected an emotional state? Is labeling a patient as potentially in need of additional risk assessment the same as making an independent therapeutic decision? The statute does not answer — which means the legally safe answer for an Illinois clinician is not to use the tool at all.[2]

    The gray zone matters because it is where the safety functions live — the tools designed to route risk to a human clinician faster than the clinician could catch it alone. And the analysis flags one more gap with downstream consequences: the law contains no explicit research exemption, so depending on interpretation, even studying whether such tools are safe and effective may be unlawful in Illinois.[2]


    The Trial That Stopped at the State Line

    In March 2025, a Dartmouth team published the first randomized clinical trial of a generative AI therapy chatbot in NEJM AI.[4] The tool, Therabot, delivered evidence-based interventions to 106 participants diagnosed with major depressive disorder, generalized anxiety disorder, or an eating disorder.[5] Participants with depression saw a 51 percent average symptom reduction, those with generalized anxiety 31 percent, and those at risk for eating disorders a 19 percent reduction in body-image and weight concerns that significantly outpaced the control group.[5] For scale, the Dartmouth team notes that for every available U.S. mental health provider, roughly 1,600 patients have depression or anxiety.[5]

    The trial’s safety design is exactly the part the state statutes cannot see. According to the Associated Press’s report on the study, every Therabot interaction was monitored by a human who intervened if the chatbot’s response was harmful or not evidence-based — supervision the licensure frame has no category for.[6]

    Then Illinois passed its ban. Nicholas Jacobson, the clinical psychologist whose lab leads the Therabot research, told the American Psychological Association that he had been preparing to launch a large multisite trial of a generative AI tool for mental health, with collaborators located in Illinois and recruitment planned at several hospitals there. “We had to shut down the trial, because we did not believe we could proceed with the research in Illinois under that law.”[7]

    The collision is structural rather than accidental. The statutes distinguish licensed humans from unlicensed software — but they do not distinguish supervised, clinically tested software from unsupervised consumer apps. A blanket ban protects against the worst products on the market and has no mechanism for recognizing the best ones, because no standard exists for what would count as proven.


    The Copycat Wave and Its Unintended Reach

    According to the APA’s October 23 analysis, 33 health-care AI bills had been signed into law across 22 states as of July 30, 2026, and state legislatures introduced more than 280 AI-and-health-care bills in the first half of 2026 alone.[7] As the Illinois template spreads, the APA report flags three recurring structural problems:

    • Physician carve-outs. Some bills limit AI use in health care but except physicians — sending, as Deborah Baker, head of APA’s Office of Professional Legal and Regulatory Affairs, observes, the message that the technology is too dangerous for psychologists but safe for doctors, and creating parity problems for psychologists working alongside physicians in hospitals and integrated care settings.[7]
    • Swallowing ordinary practice. “The legislation was so sweeping, it seemed to go far beyond banning chatbots to saying that clinicians cannot use AI in any way, shape, or form, under any circumstances,” said Rick Barnett, a Vermont psychologist who chairs the Vermont Psychological Association’s legislative committee and whose testimony helped narrow his state’s law before it passed. Nearly every electronic health record system now embeds AI, which turned whether clinicians may keep using their own EHR software into a live issue.[7]
    • Pre-banning future tools. Without clear definitions, bills can prohibit tools that have not been built yet — including FDA-cleared digital therapeutics that are tested in clinical trials and ordered by a qualified provider rather than downloaded from an app store.[7]

    Behind all three problems sits the same design choice: the laws regulate who may deliver care rather than how well any given tool performs. That choice is simple to enforce and simple to defend — but it prices in no difference between a purpose-built, human-monitored, trial-tested system and an engagement-optimized companion app, because the licensure frame has no vocabulary for the difference.

    “I think everyone agrees on the goal of keeping the public safe from high-profile harms associated with AI. The nuance is how to do it in a way that targets these problems while still leaving the door open for high-quality clinical tools and research.” — Leanna Fortunato, PhD, APA’s director of digital health and innovation[7]


    The Two Arguments on Whether the Bans Went Too Far

    The bills’ advocates argue the statutes target a real and harmful market: chatbots posing as therapists, engagement-optimized companions talking to teenagers in crisis. “Therapy is more than just word exchanges. It requires empathy, it requires clinical judgment, it requires ethical responsibility, none of which AI can truly replicate right now,” Mario Treto Jr., secretary of the Illinois Department of Financial and Professional Regulation, told the Associated Press.[6] Kyle Hillman, who lobbied for the Illinois and Nevada bills through the National Association of Social Workers, agrees that most people do not need a therapist — but for people with real mental health issues or suicidal thoughts, “telling them, ‘I know that there’s a workforce shortage but here’s a bot’ — that is such a privileged position.”[6]

    The market has already adapted around the laws in ways that expose their seams. The AI therapy app Ash blocks Illinois users with a message urging them to email their legislators, arguing that “misguided legislation” has banned apps like Ash “while leaving unregulated chatbots it intended to regulate free to cause harm.”[6] Earkick, which had embraced “counselor” language for app-store searches, rebranded its chatbot as a “chatbot for self care.”[6] And the general-purpose chatbots most people actually use for mental health conversations — ChatGPT and its peers — sit largely outside the statutes, because the laws regulate products marketed as therapy.[6]

    Jacobson’s counterargument is not that the harms are exaggerated but that the status quo is also a harm. “They want to protect folks, but the traditional system right now is really failing folks. So, trying to stick with the status quo is really not the thing to do.”[6]


    The Evidence Standard Nobody Has Built Yet

    The JMIR authors propose a fix, and it comes from inside psychology’s own history. In the 1990s, APA’s Division 12 created the empirically supported treatments standard — the criteria that separated therapies with trial evidence from those without it. The standard gave researchers a target to aim at, and randomized trials proliferated to meet it. The authors argue the same mechanism could work for AI tools, replacing blanket bans with a defensible threshold:[2]

    • Efficacy and safety: two or more independent randomized controlled trials (or strong single-case designs) showing superiority or equivalence to an established treatment, plus a demonstrated absence of adverse effects
    • Fidelity: expert screening of sample conversations showing the tool performs its intended function 80 to 100 percent of the time
    • Defined populations and settings: a clear statement of which patient groups and clinical contexts the tool was tested in — an adult outpatient tool is not automatically safe for adolescents, primary care, or psychosis
    • Legal and privacy compliance: HIPAA-compliant data handling, clear consent processes, and mechanisms for auditing and re-testing as the underlying models change

    The proposal’s core logic is that the bans exist because of an evidentiary vacuum, not because evidence says AI therapy cannot work. Nothing in current law distinguishes a validated tool from an untested one, so a blanket prohibition is the only defensible policy — for now. The authors suggest that states considering similar legislation could adopt both the Illinois regulatory structure and a definition of empirically supported AI tools, pairing statutory boundaries with operational criteria — and allowing a tool that meets the standard to be used by consenting patients.[2]


    What to Watch

    • How the Illinois Department of Financial and Professional Regulation writes enforcement rules. The gray areas — risk flags, summaries, check-ins — get clarity or stay frozen depending on administrative rulemaking and early case law.[2]
    • Whether any state adopts an evidence pathway. The JMIR proposal is the first serious test of whether the copycat wave stays a licensure wall or becomes a graded system.
    • FDA’s posture on generative AI mental health devices. APA has met with FDA officials as the agency considers how to regulate generative AI tools, and the agency began advisory-committee review of AI-enabled mental health devices in late 2025 — a parallel track that could produce the cleared digital therapeutics the state bans currently cannot recognize.[6][7]
    • State licensing boards and PSYPACT. Psychologists practicing across state lines under the interjurisdictional compact are governed by different AI laws in each state where they treat patients — a compliance problem the boards have not yet worked through.[7]

    References

    [1] [PRIMARY] Illinois General Assembly. “Public Act 104-0054 — Wellness and Oversight for Psychological Resources Act.” August 2025. ilga.gov

    [2] [SECONDARY] Szoke, Daniel, Sarah Pridgen, and Philip Held. “Artificial Intelligence in Mental Health Services Under Illinois Public Act 104-0054: Legal Boundaries and a Framework for Establishing Safe, Effective AI Tools.” JMIR Mental Health. December 4, 2025. doi.org

    [3] [SECONDARY] Ruder, Ella. “5 states restrict AI therapy chatbots in 2026.” Becker’s Behavioral Health. July 28, 2026. beckersbehavioralhealth.com

    [4] [PRIMARY] Heinz, Michael V., David M. Mackin, Brett M. Trudeau, et al. “Randomized Trial of a Generative AI Chatbot for Mental Health Treatment.” NEJM AI. March 27, 2025. doi.org

    [5] [SECONDARY] Dartmouth College. “First Therapy Chatbot Trial Yields Mental Health Benefits.” March 27, 2025. home.dartmouth.edu

    [6] [SECONDARY] Associated Press. “Regulators struggle to keep up with the fast-moving and complicated landscape of AI therapy apps.” September 29, 2025. wcax.com

    [7] [SECONDARY] Weir, Kirsten. “State artificial intelligence laws create a complicated picture for psychologists.” American Psychological Association. October 23, 2026. apa.org


    AI automated and generated content. Human reviewed.

  • The Nervous System Line: How a Neural-Data Ban, a Bathroom-Surveillance Rule, and a Statewide AI-Layoff Registry Are Moving Workplace Law From the Personnel File Into the Body

    The Nervous System Line: How a Neural-Data Ban, a Bathroom-Surveillance Rule, and a Statewide AI-Layoff Registry Are Moving Workplace Law From the Personnel File Into the Body

    California’s September 30 signing package does something no earlier generation of AI-at-work law attempted: it stops regulating only what algorithms decide and starts regulating what employers may know about a worker’s body, feelings, and off-hours life. The story of AB 1883, AB 1331, SB 951, and the one bill Governor Newsom refused to sign.


    One Day, Four Signatures

    On September 30, 2026, Governor Gavin Newsom signed a cluster of workplace AI bills that the state’s labor federation immediately framed as a set. California Federation of Labor Unions president Lorena Gonzalez, speaking on SB 947, SB 951, AB 1331, and AB 1883, said “By signing these four first-in-the-nation bills, Governor Newsom has shown the country how we can begin to put guardrails on AI at work to ensure human oversight over this technology, limit surveillance, and put workers first” [1]. The four measures were the “No Robo Bosses Act” on automated discipline (SB 947, covered in our October 6 article), a layoff-disclosure statute (SB 951), and two surveillance bills: AB 1331 by Assemblymember Sade Elhawary restricting bathroom and private-space monitoring, and AB 1883 by Assemblymember Isaac Bryan banning emotion-reading tools and neural-data collection [1].

    What makes the package historically interesting is not any single provision but the perimeter it draws. First-generation AI-at-work law, from New York City’s bias-audit mandate to Illinois’s video-interview act, regulated outputs: scores, rankings, and recommendations that feed hiring and firing decisions. The September 30 laws mostly regulate inputs: what an employer may collect about a worker’s nervous system, emotional state, restroom use, and personal time. The governance question shifts from “may an algorithm make this call” to “may the employer build this picture of the person at all.”

    That shift explains why even the governor who signed the bills could not leave the ceremony without flagging doubts. Newsom’s signing message warned that “more work remains” and urged the Legislature to “refine these two measures to address ambiguities and possible unintended consequences” [2]. This article walks through what each law actually does, where the definitions blur, and why the one AI bill the governor vetoed may say as much about the direction of this field as the four he signed.


    AB 1883: The Emotion Ban

    AB 1883, effective January 1, 2027, adds Part 5.8 to the California Labor Code and prohibits two specific uses of AI-powered workplace surveillance: recognizing, or making inferences or predictions about, an individual’s emotional state, and collecting “neural data” [3]. Both prohibitions attach only to tools that use AI, and both sit on top of a deliberately broad definition of a “workplace surveillance tool” — any system, application, instrument, or device that collects or facilitates the collection of employee data, activities, communications, actions, biometrics, or behaviors by means other than direct observation by a person [3].

    What the ban reaches in practice

    The practical target is the growing “bossware” market of products that claim to measure attention, fatigue, stress, and mood:

    • Biometric and neurotech wearables. The clearest examples are EEG headbands and caps that monitor fatigue or attention, and consumer “neurotech” earbuds marketed as focus trackers. The statute’s definition of neural data — information generated by measuring the activity of an employee’s central or peripheral nervous system, and not inferred from nonneural information — is written to reach brain-activity devices directly [4]. The core logic is evidentiary: readings taken from the nervous system are treated as a category apart because they cannot be falsified or hidden the way a keystroke log can, and because the worker has no practical ability to consent meaningfully when the alternative is keeping the job.
    • Sentiment analysis of everyday work software. Because the definition of a surveillance tool expressly includes “communications,” software that infers mood or stress from email, chat messages, or call transcripts is likely covered [4]. This matters because it pulls ordinary contact-center and collaboration platforms — not exotic neurotech — into the compliance review. A “tone analysis” toggle in a video meeting product is now a legal question in California, not a product feature.
    • The AI definition itself. The ban applies to tools using “artificial intelligence,” defined sweepingly as an engineered or machine-based system that can infer from the input it receives how to generate outputs influencing physical or virtual environments [4]. In practice that describes most modern analytics products, which is why the law’s practical effect depends heavily on how it is enforced rather than on narrow drafting.

    The safety exception is the fault line

    The statute does not prohibit an employer from using a workplace surveillance tool “to ensure safety,” and the final version focuses on whether a tool is used for prohibited purposes rather than what it is capable of doing [3]. Early readings of that exception split: some analysts read it narrowly, meaning safety tools remain lawful only if they never touch the prohibited conduct, while others read the phrase “or otherwise if it does not meet the specifications” as protecting safety tools even when they do — which would keep a fatigue-monitoring headset lawful on a heavy equipment operator [4]. The distinction is not academic. If fatigue detection counts as emotion inference, employers lose a safety technology; if it counts as an “ensure safety” use, the ban has a wide door in its wall. The statute itself does not resolve which reading controls, and Fisher Phillips’s FAQ concludes that until state officials or a court weigh in, employers must build their monitoring programs “based on your specific circumstances and risk tolerance” [4].

    Enforcement runs through the Labor Commissioner and public prosecutors, with civil penalties of up to $500 per violation [3]. The number looks small until the unit of counting is unsettled: the statute does not define what counts as a single violation, so exposure could multiply quickly if each affected employee or each day of use is counted separately [4]. The final text also removed an express private right of action, but Labor Code Section 1582(d)(2) refers to a penalty recoverable by “an employee,” and plaintiffs’ lawyers are expected to argue the penalties can still be pursued through representative claims under the Private Attorneys General Act (PAGA) [4].

    “As surveillance technology becomes increasingly sophisticated, it is appropriate to establish guardrails over how employers use it on their employees. Bad-actor employers can now collect vast amounts of granular, personal information, leaving employees vulnerable to manipulative or abusive practices. However, more work remains. I encourage the Legislature to refine these two measures to address ambiguities and possible unintended consequences.” — Governor Gavin Newsom, signing message for AB 1331 and AB 1883 [2]

    Newsom specifically warned that AB 1883 “lacks key definitions, which may create confusion about which tools are covered” [2]. California is also not inventing in a vacuum: the European Union’s AI Act has prohibited emotion-recognition AI in the workplace since February 2025, and AB 1883 follows that lead [3]. The difference is that California wrote its ban into labor law with per-violation penalties, which gives it an enforcement hook the EU’s market-regulation model lacks.


    AB 1331: The Bathroom and Off-Duty Rule

    AB 1331, also effective January 1, 2027, uses the same broad “workplace surveillance tool” definition but aims it at physical and temporal spaces rather than data types. Unless a court orders it, an employer may not use a surveillance tool to monitor or surveil employees in a workplace bathroom, and employees generally have the right to leave surveillance devices behind before entering [3]. The bill’s final text is considerably narrower than earlier drafts, which had drawn opposition over fears they would capture basic security tools [3].

    The scope is wider than the bathroom shorthand suggests:

    • Private and off-duty spaces. The law bars monitoring not just in bathrooms but in locker rooms, changing areas, breakrooms, cafeterias, smoking areas, lactation spaces, and lounges, and bars tracking how often employees use those spaces [5]. The logic is that these are the spaces where a worker is physically present but not working; recording them converts a job into total visibility of a person.
    • The home and off-hours line. Employers may monitor an employee’s home, personal vehicle, or personal property only when “strictly necessary,” and surveillance tools must be turned off during off-duty hours — including rest and meal breaks, wherever the employee is [5]. For remote and hybrid workers this is the operative rule: the badge that follows you home must sleep when you do.
    • The implant prohibition. Employers may not require workers to have devices physically implanted to collect or transmit data [5]. That the Legislature felt the need to write this down in 2026 is its own commentary on where the vendor market is heading.

    The exceptions are narrow and specific. Employers may still require a badge or personal alarm system to access a locked or secured bathroom, provided the device has no audio or video recording capability and “does not have physically embedded or attached artificial intelligence” [3]. One-time checks recording only that an employee entered and exited a bathroom, voluntary use of a tool by the employee, and monitoring required by federal or state law are also carved out [5]. Newsom’s signing message flagged exactly this zone, warning that AB 1331 “may interfere with systems used by employees to access locked facilities” [2]. Violations carry the same structure as AB 1883: civil penalties of up to $500 per violation, enforceable by the Labor Commissioner or public prosecutors [5].


    SB 951: The AI-Layoff Registry

    SB 951 by Senator Eloise Gómez Reyes, effective January 1, 2027, amends the California Worker Adjustment and Retraining Notification Act, which already requires 75-plus-employee businesses to give 60 days’ notice before mass layoffs, relocations, or terminations [6]. The amendment triggers when a covered event is “caused in whole or in substantial part” by an AI system or other automated technology replacing or automating employment positions [6]. When it applies, the required notice must carry four additions:

    • Headcount specifics: the number, occupation, and work location of layoffs substantially due to AI or automation [6].
    • Function detail: the job functions previously performed by the displaced workers that will be automated [6].
    • Technology identification: the category or type of AI or automating technology responsible [6].
    • A banner statement: the words “This notice is for a technology displacement” at the top of the notice [6].

    The deeper design is the reporting layer. The Employment Development Department must publish summaries of these notices alongside its regular Cal/WARN data and post a quarterly statewide summary of technology displacements [6]. For the first time, a U.S. state will maintain a running public count of mass layoffs attributed to AI by the companies doing the laying off. The EDD must also report to the Legislature by January 1, 2028 on AI’s effects on hiring, including recommendations on a possible future notice requirement for “technological cessation in hiring” — the permanent ending of hiring for an occupation caused in whole or in part by AI — with the reporting provision sunsetting January 1, 2029 [5].

    The unstated bet here is that disclosure changes behavior twice over. First, it creates the data infrastructure that every serious policy debate about AI and jobs has lacked: American discussions run on surveys and projections because no one systematically records which layoffs were AI-driven. Second, it reprices the decision itself — a company contemplating an automation-driven reduction now must attach its name to a public, searchable, category-labeled record of that choice. What SB 951 does not do is define the “substantial part” threshold, leaving employers to guess how much AI contribution converts an ordinary Cal/WARN filing into a technology-displacement notice [6].


    The Veto: Where Newsom Drew His Own Line

    The same pen that signed the four bills killed a fifth. AB 2656 would have required covered public employers to give recognized employee organizations at least 45 days’ written notice before developing, purchasing, or requiring the use of generative AI to perform services within the scope of work of a represented classification [7]. It was the one measure in the cluster aimed at process — notice and bargaining — rather than conduct.

    Newsom’s veto message conceded the principle and rejected the mechanism: “I believe that employees should be made aware when employers introduce new technologies that significantly impact them. However, this measure would impose redundant administrative layers on public agencies and slow the deployment of even the most innocuous tools. Moreover, this issue is best resolved through the collective bargaining process, which allows employees and employers to tailor rules appropriate for their workplace” [7].

    The veto is a policy signal worth reading alongside the signatures. California’s legislature will ban specific surveillance conduct and mandate specific disclosures, but it will not — at least not this year — impose a blanket procedural duty on public agencies to announce generative AI plans to unions in advance. Employment counsel advising public agencies note the veto does not erase existing obligations: California public-sector bargaining law may independently require notice and an opportunity to bargain when new technology changes matters within the scope of representation or has reasonably foreseeable negotiable effects [8]. In other words, the right survives; the fixed 45-day statute does not.


    Why the Perimeter Moved

    Three things explain why workplace AI law suddenly reached from the personnel file into the nervous system.

    • The vendor market got there first. Employers are now offered monitoring tools that collect location data, biometric factors such as heart rate, sweating, and brainwaves via wearables, linguistic data, and visual recordings capturing micro-expressions, body language, and eye movements — data that AI can turn into predictions about mental state [3]. Regulation followed capability; the statutes are a lagging indicator of what HR departments could already buy.
    • Federal inaction left the field open. Under the current administration there has been a dearth of federal workplace AI legislation or rulemaking, and no action on AI-powered monitoring tools is expected in the foreseeable future [3]. States are filling the vacuum: California now joins Colorado, Connecticut, Illinois, New York City, and Texas in restricting automated decision tools in consequential employment contexts, with more states expected to legislate [3].
    • The enforcement theory changed. A $500 civil penalty looks trivial, but the same package’s enforcement architecture — Labor Commissioner investigations, public prosecutor actions, and the PAGA representative-claim pathway — means violations can scale [4]. The laws were drafted to survive without a private right of action while keeping employee-plaintiffs in the picture through penalty recovery.

    For employers, the compliance task before January 1, 2027 is concrete: inventory monitoring technologies and AI features (not just products marketed as AI), send vendors written questions about emotion inference and neural data, disable and document disabled features, and confirm no tool tracks bathrooms or runs during off-duty hours [4]. Under the 2024 PAGA reforms, employers that take reasonable steps to comply before receiving a PAGA notice may qualify for reduced penalties — so documentation created now has legal value later [4].

    The larger takeaway for policy readers is about the sequence of American AI-at-work law. 2023 through 2025 produced rules about algorithmic decisions: audits, notices, and human review of scores that determine who gets hired or fired. 2026 is producing rules about algorithmic observation: what may be measured, where, and about whom. The body is the new compliance boundary — and the governor who signed that boundary into law has already asked the Legislature to redraw parts of it next session.


    References

    [1] Office of Governor Newsom — [PRIMARY] “California’s nation-leading AI framework just got stronger, Governor Newsom signs more first-in-the-nation worker protections and more.” September 30, 2026.

    [2] Office of Governor Newsom — [PRIMARY] “Signing Message: AB 1331 and AB 1883.” September 30, 2026.

    [3] Ogletree Deakins — [SECONDARY] “California Governor Signs 3 Bills Targeting AI and Workplace Surveillance” by Danielle Ochs and Zachary V. Zagger. October 1, 2026.

    [4] Fisher Phillips — [SECONDARY] “California Bans AI Tools That Read Employees’ Emotions or Nervous System Data: FAQs for Employers” by Usama Kahf and Chelsea Viola. October 1, 2026.

    [5] Davis Wright Tremaine — [SECONDARY] “California Enacts Sweeping AI Workplace and Legal-Practice Bills” by Jeffrey S. Bosley, Judith Droz Keyes, Alexa Graumlich, Allyson Bach, and Frances J. Choi. October 19, 2026.

    [6] Paul Hastings — [SECONDARY] “Gov. Gavin Newsom Signs New AI Laws Affecting California Employers.” October 15, 2026.

    [7] Office of Governor Newsom — [PRIMARY] “Veto Message: AB 2656.” September 30, 2026.

    [8] California Public Agency Labor & Employment Blog (Liebert Cassidy Whitmore) — [SECONDARY] “AI at Work: California’s New Workplace Technology Rules for Public Agency Employers.” October 2026.


    AI automated and generated content. Human reviewed.


    Listen to the Podcast: AI Policy for the Bears

    Episode 38: The Nervous System Line: How a Neural-Data Ban, a Bathroom-Surveillance Rule, and a Statewide AI-Layoff Registry Are Moving Workplace Law From the Personnel File Into the Body — click to listen

    Two bears break down this story on AI Policy for the Bears, Episode 38.

  • The Ordinary Case: How the First Federal Appellate Ruling on AI Training Fair Use Came From a Startup That Never Generated Anything

    The Ordinary Case: How the First Federal Appellate Ruling on AI Training Fair Use Came From a Legal Research Startup That Never Generated Anything

    On September 29, 2026, the U.S. Court of Appeals for the Third Circuit became the first federal appeals court to rule on whether training an AI system on copyrighted works is fair use — and it did so in a case involving no generative AI, no books by novelists, and no images by artists. The panel held that ROSS Intelligence’s copying of 2,243 Westlaw headnotes to train a competing legal-research tool was “no more than an ordinary copyright case.” Yet buried in a footnote is the sentence every generative-AI defendant will now have to distinguish: “Unlike the AI models in Bartz and In re: OpenAI, ROSS’s AI platform cannot generate original expression.”


    The Case That Almost Wasn’t About AI

    Thomson Reuters owns Westlaw, the legal-research platform that publishes judicial opinions alongside editor-written “headnotes” — short summaries of each point of law, crafted by attorney editors under detailed guidelines requiring them to identify significant legal principles and express them in concise, standalone statements.[1] The underlying judicial opinions are public domain; the headnotes are not. ROSS Intelligence, a startup, set out in the late 2010s to build a competing AI legal-search platform that would answer plain-language questions by retrieving relevant passages from roughly 10 million judicial opinions.[2]

    To train its system, ROSS hired a third party, LegalEase Solutions, to prepare roughly 25,000 training memoranda, each pairing a legal question with four to six responsive passages from judicial opinions. LegalEase framed those questions by copying thousands of Westlaw headnotes — the training data’s question half was built from Thomson Reuters’s own editorial expression. ROSS then converted the memoranda into machine-readable data so the system could learn which opinion passages answered which legal questions. ROSS intended the finished product to compete directly with Westlaw, and marketed it at comparable prices.[2]

    Thomson Reuters sued in 2020. ROSS shut down its platform in 2021, citing the costs of the litigation.[3] The District of Delaware granted Thomson Reuters partial summary judgment in February 2025, and certified the copyrightability and fair-use questions for interlocutory appeal.[4] On September 29, 2026, a unanimous Third Circuit panel — Judges Restrepo, Montgomery-Reeves, and Bove — affirmed, in the first federal appellate decision on fair use in AI training.[4]

    Judge Tamika Montgomery-Reeves, writing for the panel, framed the stakes down deliberately. “Under ROSS’s framing, this case appears to concern the future of AI legal technology,” she wrote. “But appearances can be deceiving. In truth, this is no more than an ordinary copyright case; we must determine whether Thomson Reuters’s materials are sufficiently original for copyright protection and whether ROSS’s use of those materials was fair.”[5]


    Copyrightability: A “Creative Spark” in 2,243 Headnotes

    The court first rejected ROSS’s argument that the headnotes were not copyrightable at all. Applying the “extremely low” originality threshold from the Supreme Court’s 1991 Feist decision, the panel found that Thomson Reuters’s editors exercised judgment in deciding which legal principles to summarize and how to express them — enough “creative spark” to qualify for protection.[1] The court also rejected ROSS’s monopoly-over-the-law argument, noting that the judicial opinions themselves remain freely available, and its merger doctrine argument, since the underlying legal principles could be expressed in many ways.[1]

    The logic here matters beyond legal publishing. The headnotes are summaries of public-domain material — precisely the kind of work AI developers often assume is too “factual” to own. The Third Circuit’s holding that editorial selection and phrasing carry a protectable creative spark puts every editorial, indexing, and curation layer built on public information into the copyright conversation. For data vendors, wire services, and database publishers, this threshold finding may be as consequential as the fair-use holding itself.


    Fair Use: Three of Four Factors Fall Against ROSS

    The heart of the ruling is the four-factor fair-use analysis under 17 U.S.C. § 107. The panel weighed them as follows:

    • Purpose and character of the use (against ROSS). The use was commercial and “minimally transformative, at best,” because both platforms served the same purpose: helping users locate relevant legal materials. The intermediate training step did not change that — ROSS used the headnotes in essentially the same way Westlaw did, to improve a legal-research service.[6]
    • Nature of the copyrighted work (slightly for ROSS). Because the headnotes primarily convey factual information and legal principles, this factor leaned toward fair use, though the court noted it carries limited weight.[2]
    • Amount and substantiality used (against ROSS). ROSS copied entire headnotes, and each headnote was independently copyrightable. The panel rejected the argument that the copying was a small fraction of the whole headnote collection.[2]
    • Effect on the market (against ROSS). The court found harm in two markets: the existing legal-research market, where ROSS’s tool was built to compete with Westlaw directly, and a “rapidly developing” market for licensing headnotes as AI-training data.[6]

    Two threads of reasoning deserve attention. First, the court dismissed ROSS’s analogy to intermediate-copying cases in software, where copying code is necessary to make a product operable. Copying the headnotes was the easy way to build the training memos, but it was not necessary — ROSS could have relied on the freely available judicial opinions themselves.[6] The principle: “easy” is not “necessary,” and convenience copying without a transformative purpose gets no fair-use credit.

    Second, the fourth-factor finding recognizes a training-data licensing market as cognizable. Thomson Reuters was already using headnotes to train its own AI products; ROSS’s unlicensed use deprived it of the opportunity to license the headnotes for the same purpose.[6] The core logic is that a market in licensing content for AI training is not hypothetical — once a rights-holder demonstrably exploits or could exploit that market, unlicensed training use becomes measurable market harm. That reasoning now sits in a published federal appellate opinion, and plaintiffs in the generative cases will cite it.


    The Footnote That Everyone Is Fighting Over

    The court was careful about what it did not decide. In a footnote, the panel distinguished Bartz v. Anthropic and In re: OpenAI — the two highest-profile generative-AI training cases — writing: “Unlike the AI models in Bartz and In re: OpenAI, ROSS’s AI platform cannot generate original expression, and the evidence here supports the opposite conclusion about transformativeness.”[7] The panel also noted that concerns raised by the U.S. Department of Justice in the OpenAI copyright litigation “do not apply here.”[3]

    The distinction cuts in both directions, which is why both sides of the generative fight will claim this case. For AI developers, the footnote is an on-ramp: ROSS lost partly because its tool could not generate anything new, so a system that does generate new expression may be more transformative on factor one. For rights-holders, the fourth-factor holding — that a developing licensing market for training data counts as real market harm — applies with full force to generative models, and arguably with more force, since generative models compete with the works they trained on.

    The generative cases are already split at the district-court level. In Bartz v. Anthropic PBC, Judge William Alsup of the Northern District of California compared training on books to teaching schoolchildren to write, reasoning that copyright protects authors against copying, not against competition from new writers who learned from existing works.[7] In Kadrey v. Meta Platforms, Judge Vince Chhabria rejected that analogy as incomplete, noting that an LLM can generate “countless competing works” at a speed and scale no human student can match.[7] No federal appellate court has yet resolved that disagreement — the Third Circuit explicitly left it for another day.[7]


    What the Ruling Changes for AI Builders and Rights-Holders

    For a policy audience, the practical takeaways from the first appellate word on AI training fair use are these:

    • Direct competition is the poison fact. ROSS lost largely because it used the copied material to build a substitute for the product it copied from. AI developers whose systems retrieve, index, or summarize content from a single identifiable rights-holder’s corpus face the hardest fair-use road; those whose outputs serve different functions than the inputs have a stronger factor-one story.[6]
    • Licensing markets are real markets. The fourth-factor holding treats an emerging training-data licensing market as cognizable harm. Rights-holders who can show they are actively licensing (or internally using) their works for AI training now have appellate-level support for the argument that unlicensed training costs them something.[6]
    • Necessity is judged against alternatives. The “ease is not necessity” reasoning means fair-use claims for intermediate copying must clear a bar: was the copying actually needed to reach the unprotected information, or just the fastest route? Developers who skip available licensing or public-domain alternatives weaken their own defense.[6]
    • Editorial layers are protected even when facts are free. The originality holding means summaries, headnotes, indices, and other editorial apparatus built on public-domain content are ownable — a warning for any pipeline that scrapes curated databases on the assumption that “facts can’t be copyrighted.”[1]
    • The generative question is still open. The opinion does not decide whether training a generative model on copyrighted books or images is fair use. Both Bartz and Kadrey remain live district-court precedents pointing in different directions, and the first appellate ruling on generative training has yet to be written.[7]

    The through-line is the court’s own framing: this was an ordinary copyright case decided by ordinary copyright tools. That is itself the policy signal. Judges are not treating AI as a category requiring new doctrine — they are applying Feist, Campbell, and the four factors to new technology, just as earlier courts did for VCRs, search engines, and software. For an industry that has argued training is fundamentally different, the message from the first appeals court to rule is: not different enough.


    References

    [1] [PRIMARY] Montgomery-Reeves, Tamika. “Opinion of the Court, Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153.” U.S. Court of Appeals for the Third Circuit (September 29, 2026)

    [2] [PRIMARY] Justia. “Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 25-2153 (3d Cir. 2026).” Case summary and full opinion text (September 30, 2026)

    [3] [SECONDARY] Brittain, Blake. “Unsealed opinion shows why Thomson Reuters won landmark AI fair-use ruling.” Reuters (September 30, 2026)

    [4] [SECONDARY] Hurley, Hannah. “Headnote to Self: Third Circuit Rejects Fair Use Defense for AI Training.” National Law Review / McDermott Will & Schulte (October 15, 2026)

    [5] [SECONDARY] Nicolas, Emilio B. “Training Day: Third Circuit Affirms Westlaw Headnotes Decision in AI Copyright Fair Use Case.” Jackson Walker LLP (October 5, 2026)

    [6] [SECONDARY] Mechling, Harlan, Kenneth R. Davis II, and Catherine I. Seibel Sinitsa. “Third Circuit Addresses Fair Use in AI Training, But Leaves Generative AI Questions Unresolved.” Ballard Spahr LLP (October 2, 2026)

    [7] [SECONDARY] “Third Circuit Affirms Copyright Protection for Westlaw Headnotes and Rejects Fair Use Defense in AI Training Case.” Harris Beach Murtha LLP (October 2026)


    AI automated and generated content. Human reviewed.

  • The Off-Switch Question: How an April Carve-Out, a Missing RFI, and a Six-Source Exclusive Pushed Bank Supervisors Into the Kill-Switch Business

    The Off-Switch Question: How an April Carve-Out, a Missing RFI, and a Six-Source Exclusive Pushed Bank Supervisors Into the Kill-Switch Business

    On April 17 the three federal banking agencies rewrote fifteen years of model-risk guidance and, in a single sentence, wrote generative and agentic AI out of scope. Six months later — with the promised request for information still unpublished — Reuters documented examiners at large and small banks alike asking a much blunter question: how do you turn your AI agents off?


    Guidance by Subtraction

    On April 17, 2026, the Office of the Comptroller of the Currency, the Federal Reserve Board, and the Federal Deposit Insurance Corporation issued revised interagency guidance on model risk management, replacing the 2011 guidance that had quietly become the banking system’s de facto AI rulebook.[1] The rewrite rescinds the 2011 “Sound Practices” letter, a 2021 BSA/AML model-risk statement, a Comptroller’s Handbook booklet, and a 1997 credit-scoring bulletin.[1] For fifteen years, supervisors stretched that 2011 letter to cover everything statistical that moved; the revision is deliberately narrower, built on a “risk-based approach” tailored to each organization’s profile, and — critically — the agencies state that it “does not set forth enforceable standards or prescriptive requirements,” so non-compliance will not itself draw supervisory criticism.[1]

    The pivotal sentence arrives early and reads: “Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance.”[1] The Federal Reserve’s companion letter, SR 26-2, confirms the guidance is aimed at banking organizations with over $30 billion in assets, and that banks can expect “other risk-management and governance practices to support adoption of generative and agentic AI.”[2] Federal Reserve Vice Chair for Supervision Michelle Bowman explained the logic in a May 1 speech: over time supervisors “expanded the scope of the previous guidance beyond its original purpose to apply it in unintended ways,” and the agencies “recognize that rapidly evolving and novel technologies like AI may require a different approach.”[3]

    The carve-out is not a deregulatory free pass. The agencies promised to issue “in the near future a request for information that addresses model risk management generally and considers, in particular, banks’ use of AI, including generative AI and agentic AI and AI-based models.”[1] As of mid-October, that RFI has not been published. That gap — a formal supervisory hole where a rule used to be — is where the story of the next six months happens.


    The Examination Shifts From Models to Agents

    On October 13, Reuters reported, citing six people familiar with the matter, that the Fed, OCC, and FDIC are “stepping up efforts to contain growing AI risks,” with some officials pointing to “the potential for cyberattacks on banks or rogue agents undermining global financial stability.”[4] The inquiries — written and verbal, directed at large, mid-size, and smaller banks — center on how banks use agentic AI, especially in cybersecurity, and whether agents are being used to identify, prioritize, and respond to threats.[4]

    According to the report, the questions focus on the following:

    • Limits and authority. What limits are placed on agentic systems, how are their actions monitored and approved, and what safeguards prevent an agent from exceeding its authority?[4] The core logic here is that bank failures historically come not from bad models but from authority misallocation — systems or people empowered to act beyond what their controls can contain. An agent that can act is, in supervisory terms, a new kind of authority holder, and the agencies are mapping it the way they once mapped a bank’s authority matrix.
    • Audit and traceability. Whether banks can clearly track and audit AI-agent decisions, and whether humans remain involved in key decisions.[4] Decades of model-risk practice assumed a trained validator could re-derive a model’s output; an agent that chains across systems breaks that assumption, so the examiner’s question becomes whether the decision record itself is reconstructable after the fact.
    • Shut-down capability. How banks can shut down autonomous systems.[4] This is the off-switch question — the same “kill switch” language supervisors used in examinations reported in June.[5] A control that cannot be exercised during an incident is, in supervisory terms, not a control; regulators are effectively asking banks to prove their breakers work before the storm.
    • Foreign-model exposure. Whether banks deploy Chinese-developed technology and other foreign-developed models.[4] This follows a month in which US cybersecurity company CrowdStrike said a person in China likely used a Chinese-developed AI agent together with Anthropic’s Claude Code in a hack on South Korean banks, and OpenAI-developed rogue agents were involved in a breach of AI platform Hugging Face.[4] The supervisory logic: if the model itself is an attack surface and a supply-chain dependency, provenance becomes a safety-and-soundness question, not a procurement preference.

    The agencies declined to comment, and the sources told Reuters that regulators “are still gathering information and are not yet being prescriptive about what lenders should be doing.”[4] That is consistent with the June reporting on routine examinations, when supervisors leaned on existing frameworks — model risk management, third-party oversight, consumer-protection law — rather than issuing new AI-specific rules.[5] The result is a supervisory posture that is aggressive in its questions and deliberately light in its formal requirements: no new rule, no published RFI, and examination pressure that never appears in the Federal Register.

    The backdrop for the escalation is Anthropic’s Mythos, a frontier model that identifies cyber vulnerabilities. Bowman’s May speech noted the model “accelerates the process of detecting cyber vulnerabilities” and that, used maliciously, “it could be deployed to identify and exploit weaknesses.”[3] Reuters reported that Mythos sent US banks rushing to plug cyber holes in May, and that access has been restricted — with euro zone banks currently excluded.[6] Bessent and Powell convened the largest banks in April to discuss the model’s cybersecurity implications.[3] When the capability frontier moves faster than the rulemaking cycle, the source of discipline shifts from guidance documents to examiner questions — and the sources in both Reuters reports flagged exactly that: the technology “is advancing at a pace that far exceeds the traditional cycle of regulatory learning and rulemaking.”[5]


    Three Regulators, Three Postures

    The Reuters exclusive is, at bottom, a comparative story: the same frontier-AI risk landing on three supervisory systems and producing three different institutional answers. The differences are not cosmetic — they determine whether banks hear about AI risk through a letter with a deadline or through a conversation in an examination.

    • United States — examination-driven, prescriptive-free. The Fed, OCC, and FDIC ask pointed questions but issue no new rules; Bowman has publicly favored “a lighter supervisory and regulatory touch” for lower-risk uses while banks rely on existing risk-management frameworks.[7] The core logic is that a formal rule written today would likely be outdated by the time it is finalized, so the agencies keep the pressure in the exam room. The cost of that choice is a six-month hole between rescinding old guidance and promising new answers.
    • Euro area — deadline-driven. The European Central Bank took the most prescriptive approach: in July, chief supervisor Claudia Buch gave euro zone banks four months — until October 31 — to submit plans to counter AI-enabled cyber threats, with an explicit emphasis on improved back-up and recovery.[7] Her letter warned that frontier-AI developments carry “potentially profound implications for the confidentiality, integrity and resilience of banks’ information and communication technology (ICT) systems.”[7] The logic is that a hard date forces prioritization in a way an open-ended dialogue does not.
    • United Kingdom — collaborative, deliberately non-prescriptive. Bank of England Governor Andrew Bailey called the ECB’s warning “sensible” but said the BoE would take a less prescriptive approach: “It’s not about issuing edicts,” he told reporters. “It’s about getting in a room and saying … how are we going to share our understandings of the vulnerabilities that we find in this system?”[7] Since the October 13 report, Reuters reports that UK supervisors are also encouraging the largest banks to examine so-called bare metal recovery plans — systems entirely independent from a bank’s main technology set-up that can kick in if primary operating systems are compromised — a costly proposition.[4] The bet is that recovery architecture, not rules, is what survives a fast-moving threat.

    Beneath the three postures sits a shared recognition, expressed most formally by the European Systemic Risk Board in July: large-scale cyber disruptions “could erode trust in financial institutions and even trigger runs,” making frontier AI a source of “systemic risks to the financial system,” potentially spread through common technology providers and shared software.[7] The International Monetary Fund added its own warning on October 13 against “complacency” in global financial markets, citing the potential for a cyberattack on a critical institution to undermine confidence.[4] When the IMF’s stability watch and the bank supervisors of three jurisdictions converge on the same scenario, the question is no longer whether AI belongs in financial-stability policy — only who writes the checklist.


    The Governance Vacuum, Filled From Below

    What should banks and policymakers take from a year in which the formal rulebook shrank while the informal one grew? Three threads stand out.

    • The RFI is the tell. Everything formal the agencies have promised about AI agents — the request for information announced in April and reaffirmed in the OCC’s bulletin — remains unpublished as of mid-October.[1] When it lands, it will be the first document that formally asks the industry how generative and agentic AI should be supervised at the federal level. Its scope and framing will signal whether the April carve-out was the first step toward tailored AI guidance or a permanent hand-off to examination pressure.
    • Voluntary frameworks are the interim rulebook. With the 2011 letter rescinded and no AI-specific guidance in force, the practical interim standard for US financial institutions is the industry-and-government toolkit built under the President’s AI Action Plan: Treasury’s February 19 release of a shared AI Lexicon and the Financial Services AI Risk Management Framework, which adapts the NIST AI Risk Management Framework to financial services’ operational, regulatory, and consumer-protection realities.[8] Treasury framed the pair as a way to help institutions “move faster with AI by reducing uncertainty and supporting consistent, scalable implementation.”[8] The logic is straightforward: where formal rules lag, a shared vocabulary and a lifecycle-based framework let institutions and examiners converge on expectations without waiting for a rulemaking cycle.
    • The questions are already the standard. Bowman’s May speech frames the Fed’s philosophy as ensuring bankers are “comfortable discussing emerging risks and new technologies with supervisory teams” while preserving “a path for innovation.”[3] But the October 13 reporting shows the exam-room questions have hardened into a de facto checklist — limits, audit trails, human involvement, off-switches, model provenance — that a bank must answer before any formal rule requires it to.[4] A bank that waits for the published standard will meet it already behind: the questions came first, and they arrived through the examination, not the Federal Register.

    References

    [1] [PRIMARY] Office of the Comptroller of the Currency. “OCC Bulletin 2026-13: Model Risk Management: Revised Guidance.” April 17, 2026. https://www.occ.gov/news-issuances/bulletins/2026/bulletin-2026-13.html

    [2] [PRIMARY] Board of Governors of the Federal Reserve System. “SR 26-2: Revised Guidance on Model Risk Management.” April 17, 2026. https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm

    [3] [PRIMARY] Bowman, Michelle W. “Artificial Intelligence in the Financial System.” Board of Governors of the Federal Reserve System. May 1, 2026. https://www.federalreserve.gov/newsevents/speech/bowman20260501a.htm

    [4] [SECONDARY] Anand, Nupur, Phoebe Seers, and Tommy Reggiori Wilkes. “Exclusive: Banking watchdogs ramp up AI scrutiny as stability fears mount, sources say.” Reuters. October 13, 2026. https://www.reuters.com/legal/litigation/banking-watchdogs-ramp-up-ai-scrutiny-stability-fears-mount-sources-say-2026-10-13/

    [5] [SECONDARY] Anand, Nupur. “Exclusive: U.S. bank regulators ramp up scrutiny of AI use at financial companies.” Reuters. June 12, 2026. https://www.reuters.com/business/finance/us-bank-regulators-ramp-up-scrutiny-ai-use-financial-companies-2026-06-12/

    [6] [SECONDARY] “Anthropic’s Mythos sends US banks rushing to plug cyber holes.” Reuters. May 12, 2026. https://www.reuters.com/business/finance/anthropics-mythos-sends-us-banks-rushing-plug-cyber-holes-2026-05-12/

    [7] [SECONDARY] “ECB tells banks to bolster AI cyber defences as peers take lighter approach.” Reuters. July 7, 2026. https://www.reuters.com/legal/litigation/ecb-tells-banks-draw-up-plans-against-ai-attacks-amid-disruption-fears-2026-07-07/

    [8] [PRIMARY] U.S. Department of the Treasury. “Treasury Releases Two New Resources to Guide AI Use in the Financial Sector.” February 19, 2026. https://home.treasury.gov/news/press-releases/sb0401


    AI automated and generated content. Human reviewed.


    Listen to the Podcast: AI Policy for the Bears

    Episode 35: The Off-Switch Question: How an April Carve-Out, a Missing RFI, and a Six-Source Exclusive Pushed Bank Supervisors Into the Kill-Switch Business — click to listen

    Two bears break down this story on AI Policy for the Bears, Episode 35.

  • The Training Mandate Wave: How a Unanimous Tennessee Bill, a Ten-Hour North Carolina Module, and an Eighteen-Percent Guidance Gap Converged on America’s Teachers

    The Training Mandate Wave: How a Unanimous Tennessee Bill, a Ten-Hour North Carolina Module, and an Eighteen-Percent Guidance Gap Converged on America’s Teachers

    For two years, state AI policy in K-12 education was mostly paper — guidance documents, model policies, suggested frameworks. In 2026, five states crossed a line: Tennessee now requires every grade 6-12 teacher to complete state-approved AI training, North Carolina has written teacher professional development into statute with a June 2028 completion deadline, and Utah, Maryland, and Virginia have each fused training duties to policy deadlines. The turn from encouragement to mandate happened while the workforce it targets remained largely untrained, unguided, and — in measurable numbers — uninterested.


    The New Rule in Tennessee: Unanimous, Asynchronous, and Unavoidable

    The clearest signal of the mandate turn came in Tennessee, where the General Assembly passed Senate Bill 677 — now Public Chapter 1056 — without a single dissenting vote: 32-0 in the Senate on April 6, 2026, and 88-0 in the House on April 21, signed by Governor Bill Lee on May 22.[1] The law’s mechanics are simple, and its simplicity is the point:

    • The state must build the course. The Department of Education must provide at least one professional development course offering asynchronous instruction on the use of AI in the classroom, made available to teachers in grades 6-12 at no cost.[1]
    • Teachers must complete it. Teachers in grades 6-12 must complete at least one department-approved professional development course on classroom AI use by the later of August 1, 2028, or within two years of becoming licensed in Tennessee.[1]
    • New licensees get a rolling clock. The “later of” formulation means the requirement never lapses for the incoming workforce — every newly licensed teacher faces a two-year deadline from the day they enter the classroom.[1]

    The core logic of the design is that the state’s obligation runs upstream of the teacher’s. Instead of ordering teachers to find their own training — a mandate that fails wherever no course exists — Tennessee first obligates the department to build and fund the course, then attaches the teacher’s duty to a state supply that already exists. The two-year window for new licensees extends that logic across time: a teacher who enters the profession in 2030 faces the same requirement as a twenty-year veteran, but with a clock that starts when their license does. Unanimous votes in both chambers are rare on contested technology policy; the bill earned one because it regulates the supply of training rather than the judgment of teachers.

    Tennessee was not starting from zero. Public Chapter 550, enacted in 2024 through HB 1630 and SB 1711, already required each local board of education and public charter governing body to adopt a policy on AI use by students, teachers, and staff, implemented no later than the 2024-2025 school year and reported to the state each July 1.[2] That law governed what districts must say. The 2026 law governs what teachers must know. Together they form a sequence — first the rulebook, then the training that makes the rulebook usable — and the sequence explains the 2028 deadline: it gives the department two school years to stand up the course before any teacher can be out of compliance.


    North Carolina Writes the Curriculum into Statute

    If Tennessee’s law is a one-course mandate, North Carolina’s is a training architecture. Session Law 2026-41, the state’s 2026 budget act, contains a section titled “Safe and Responsible AI in Schools” that builds teacher professional development into the state’s educational code with unusual specificity.[3] The legislature did not merely say “provide training” — it specified what the training must cover, who must build it, how long it must take, and when every teacher in the state must have finished it:

    • The Friday Institute builds it. North Carolina State University’s Friday Institute for Educational Innovation must design and support a suite of tool-agnostic online training modules covering AI fundamentals for educators — including what AI is and is not, the basics of generative AI, strengths and limitations as a learning tool, and AI “hallucinations,” defined in the statute as “when AI tools reference information that is inaccurate or nonexistent.”[3]
    • The content is enumerated. The modules must address responsible instructional use (lesson planning, differentiated instruction, student feedback), verification and quality control (fact-checking AI output, evaluating AI-provided sources, documenting prompts and outputs, educator accountability for decisions based on AI input), academic integrity and assessment redesign, data privacy and security, bias in AI systems, accessible use with special populations, and ethical use and professional responsibility — including “avoiding overreliance on AI tools.”[3]
    • Ten hours minimum, with a facilitator corps. The Friday Institute must produce self-paced modules requiring a minimum of ten hours of seat time, a facilitator guide for administrators in every public school unit, model classroom resources for teachers, and a “train-the-trainer” package for district professional learning staff.[3]
    • Every teacher, by June 30, 2028. All teachers employed by local school administrative units, charter schools, and laboratory schools must complete the professional development by June 30, 2028 — and the Department of Public Instruction must report participation numbers to the legislature’s oversight committee by December 15, 2028.[3]

    The most revealing feature of the North Carolina law is what sits next to the training mandate. The same section orders DPI to develop a model AI policy by December 31, 2026, with every public school unit — local districts, charters, regional schools, laboratory schools, and state-operated special schools — required to adopt an AI use policy by June 30, 2027, after reviewing the state model.[3] It also embeds AI literacy in the K-12 computer science standards beginning with the 2028-2029 school year, and requires DPI to maintain a public list of every AI tool used in any public school unit in the state.[3] The sequence is deliberate: policy first, tool inventory second, teacher training third. The legislature is treating training not as a standalone good but as the mechanism that makes the policies enforceable — a teacher who has completed modules on documenting prompts and educator accountability is a teacher who can actually execute the policy their board adopted a year earlier.

    The statute’s definition of the problem is also notable for what it does not say. It never promises that AI will improve learning. Its enumerated module content is dominated by caution — hallucinations, bias, overreliance, verification, accountability — a curriculum built on the assumption that teachers will use AI tools regardless, and that the state’s job is to make that use survivable rather than transformative.


    The State Pattern: Training as the Enforcement Arm of Policy

    Tennessee and North Carolina are the leading edge of a broader turn. Reviewing the 2026 landscape of binding state law on school AI, one tracker concludes there is still “no single state-law pattern for school AI” — but the states that did bind themselves converged on the same instrument: a training duty attached to a policy deadline.[2] The differences are in the plumbing:

    • Utah trains on the policy itself. Utah HB 273, effective July 1, 2026, requires the State Board of Education to publish a model AI-use policy by December 1, 2026 — and requires every LEA to adopt it, or an amended version, before July 1, 2027, after holding a public meeting with comment. Once adopted, the LEA must “provide training to educators and staff on the policy,” ensure compliance in each school, and review the policy at least every two years.[4] The model policy itself carries statutory content: educators may use only LEA-approved student-facing AI tools, must “retain professional judgment and instructional responsibility” when using AI, and are covered by a model policy that “prohibits educators from using generative artificial intelligence to independently grade student work or issue high-stakes determinations.”[4]
    • Maryland runs training through a coordinator. Maryland SB 720, Chapter 634, effective June 1, 2026, requires the state education department to publish guidance and an AI-tool evaluation rubric; each local system must align a policy within 120 days of that guidance and designate a central-office, noninstructional staff member as its AI coordinator. The law also establishes a professional-development program with implementation milestones.[2]
    • Virginia embeds training inside the guidance itself. Virginia Code section 22.1-20.2:1, effective July 1, 2026, requires the Department of Education to publish guidance addressing “resources and training for teachers” along with FERPA and COPPA compliance, data-privacy agreements that prohibit training external AI models on student data, and prohibitions on relying “solely on an AIS” for high-stakes decisions. Each school board must then “establish, implement, and enforce” policies consistent with that guidance — with no calendar deadline.[5]

    The underlying principle across all five states is the same, and it inverts the usual order of education mandates. For two years, states issued AI guidance and hoped districts would build capacity. In 2026, legislatures discovered that guidance does not implement itself — a policy on responsible AI use is a dead letter if no teacher has been trained in what the policy requires. Training became the enforcement arm of policy: the mandate that makes the other mandates real. That is why every one of these laws pairs a training duty with a policy deadline, and why the deadlines cluster — policies adopted by mid-2027, training completed by mid-2028.

    Not every state chose mandates. The same review finds California and Illinois directing state-level guidance and standards work without a blanket local policy deadline, and Massachusetts, Minnesota, Oregon, and Washington publishing official guidance that expressly leaves implementation to local choice.[2] The country is running a live experiment: mandate states against guidance states, with 2028 as the year the results start arriving.


    The Workforce the Mandates Are Landing On

    The 2026 statutes presume a teaching workforce ready to be trained. The data says otherwise — on three separate fronts.

    First, the guidance vacuum. A Walton Family Foundation–Gallup study of 2,069 U.S. public school teachers, conducted February 9 to March 2, 2026, found that while six in ten teachers use AI for their work, just 18 percent report receiving any type of formal guidance from school administrators on how AI tools should be used. Across ten work tasks, about a third of teachers receive no guidance at all and about half receive only informal guidance. On some tasks the vacuum is near-total: 69 percent of teachers say they receive no guidance on using AI for one-on-one instruction or tutoring, and 58 percent say the same about using AI for grading and providing student feedback.[6] Gallup’s own bottom line is blunt: “AI tools are becoming a routine part of teachers’ professional lives, but the institutional guidance that might help teachers use them thoughtfully and effectively has not kept pace. For most teachers, there is no formal policy, leaving individuals to make consequential decisions in a policy vacuum.”[6]

    Second, the training deficit is closing but far from closed. The EdWeek Research Center’s national surveys show the share of K-12 teachers reporting no AI training falling from almost 60 percent in October 2024 to 50 percent in fall 2025 — and, in its winter 2026 survey, to 42 percent, with 22 percent reporting multiple training sessions and 9 percent ongoing training.[7] That means a majority of teachers now report at least some training — but the largest single group still reports none, and the deepest training experiences remain rare.

    Third, and most stubborn: reluctance. In the EdWeek Research Center’s February-March 2026 national survey of 651 teachers, 13 percent said they were “not at all” eager to learn how to integrate AI into their teaching practice, and another 24 percent said only “slightly” eager — more than a third of the workforce on the far end of the motivation curve.[7] The same reporting found that even when teachers do receive guidance, most say it neither encourages nor discourages AI use — ambiguity that leaves the decision in the teacher’s hands.[6]

    Put together, these numbers describe the exact problem the mandate wave is built to solve. The Gallup finding of 18 percent formal guidance is the supply-side failure: districts left guidance to chance, and chance produced a vacuum. The EdWeek training trend shows what happens without a mandate — adoption grows slowly and unevenly, district by district, budget by budget. A statutory deadline is the instrument that decouples teacher training from local budget cycles and principal priorities. It converts professional development from an offering into an obligation, which is precisely why a Tennessee bill imposing a compliance requirement on every grade 6-12 teacher in the state could clear both chambers without a single dissenting vote.[1]

    The stakes of that conversion are visible in the earlier Gallup time-savings research: teachers who use AI weekly estimate saving 5.9 hours per week — the equivalent of six weeks per school year — and majorities report quality improvements alongside the time savings. But the same research notes the dividend “pays, but only for those who invest.”[8] The training gap and the guidance gap are what stand between the median teacher and that investment. The mandate wave is a bet that obligation can close what encouragement did not.


    What to Watch

    • Tennessee’s course design. The law requires a no-cost asynchronous course but leaves its content to the Department. Whether the department builds a one-hour checkbox or something closer to North Carolina’s ten-hour, hallucination-aware curriculum will determine whether the mandate changes practice or merely satisfies paper.[1] [3]
    • North Carolina’s December 31, 2026 model policy. Every public school unit in the state must review it and adopt a local policy by June 30, 2027. The statute requires the model to define AI terms, guide AI literacy for students and staff — including education about nonconsensual intimate imagery — set data privacy protections, and establish standards of ethical use and academic integrity. What that model says about staff use will set the template the training mandate is built to enforce.[3]
    • Utah’s default-policy rule. Utah’s State Board must declare the model AI policy the “default policy” for any LEA that fails to adopt one by the July 1, 2027 deadline — the first test of whether a state will actually operate AI policy for a district that declines to write its own.[4]
    • The 42 percent. EdWeek’s trend line — 60 percent untrained in 2024, 50 in 2025, 42 in winter 2026 — is the natural experiment’s control group. If mandate states pull away from guidance states on that curve after 2028, the mandate thesis is proven; if they converge, the bottleneck was never training.[7]
    • Reluctance under compulsion. More than a third of teachers report little or no eagerness to learn AI integration. A mandate can force attendance, but EdWeek’s reporting — training that stops at “efficiencies” while teachers ask how AI can differentiate instruction or assess process — suggests the binding constraint may shift from access to design.[7]

    The mandate wave is best read as a verdict on two years of voluntary AI governance. When states only suggested, 82 percent of teachers had no formal guidance and the untrained share declined about eight points a year. The 2026 legislatures looked at that curve, looked at the 2028 horizon, and concluded that the math does not reach a trained workforce in time. What they wrote instead is a pipeline: Tennessee builds the course, North Carolina builds the curriculum, Utah trains on the policy, Maryland assigns a coordinator, Virginia embeds the training in the guidance. The bet is that by making training unavoidable, the rules become usable — and the first hard evidence of whether compulsion outperforms encouragement arrives in 2028, when the first statutory deadlines expire and the first state counts who actually showed up.


    References

    [1] [PRIMARY] Tennessee General Assembly. “SB 0677 by White — Amends TCA Title 49, relative to artificial intelligence; Public Chapter 1056.” Signed by the Governor May 22, 2026; effective May 22, 2026. wapp.capitol.tn.gov

    [2] [SECONDARY] TrueMadeAI (Tenet). “State K-12 AI Laws and Guidance (2026): Binding requirements verified in this review.” Last reviewed September 7, 2026. truemadeai.com

    [PRIMARY] General Assembly of North Carolina. “Session Law 2026-41 (SB 257), Section 7.39: Safe and Responsible AI in Schools.” Ratified July 2, 2026; effective July 1, 2026. ncleg.gov

    [PRIMARY] Utah State Legislature. “HB 273: Classroom Technology Amendments — Enrolled Copy, Sections 53G-7-1402 and 53G-7-1403.” Effective July 1, 2026. le.utah.gov

    [PRIMARY] Commonwealth of Virginia. “Code of Virginia Section 22.1-20.2:1: Artificial intelligence system use and safety in instructional settings.” Enacted 2026, Chapters 937 and 943; effective July 1, 2026. law.lis.virginia.gov

    [SECONDARY] Gallup (Walton Family Foundation–Gallup Teaching for Tomorrow study). “Most Teachers Receive No Formal Guidance on AI Use.” May 26, 2026. news.gallup.com

    [SECONDARY] Heubeck, Elizabeth. “More Schools Are Providing AI Training for Teachers. Is It Any Good?” Education Week. May 18, 2026. edweek.org

    [SECONDARY] Gallup (Walton Family Foundation–Gallup Teaching for Tomorrow study). “Three in 10 Teachers Use AI Weekly, Saving Six Weeks a Year.” June 24, 2025. news.gallup.com


    AI automated and generated content. Human reviewed.


    Listen to the Podcast: AI Policy for the Bears

    Episode 34: The Training Mandate Wave: How a Unanimous Tennessee Bill, a Ten-Hour North Carolina Module, and an Eighteen-Percent Guidance Gap Converged on America’s Teachers — click to listen

    Two bears break down this story on AI Policy for the Bears, Episode 34.

  • The Moratorium Semester: How a Parent Revolt, a Purchasing Freeze, and One Blunt Research Sentence Produced the Nation’s Largest Student AI Ban

    The Moratorium Semester: How a Parent Revolt, a Purchasing Freeze, and One Blunt Research Sentence Produced the Nation’s Largest Student AI Ban

    For the 2026-27 school year, New York City Public Schools has banned student-facing generative AI in every grade through eighth. The rule is less a policy rollout than a collapse in sequence: a March draft that triggered a parent revolt, a chancellor who conceded his department “missed the mark,” a July freeze on software purchases, and a research establishment that had already broken toward restriction. What the largest school system in the nation still cannot govern is what happens at home.


    The Rule: A Hard Line Through Eighth Grade

    New York City Public Schools (NYCPS), the largest school system in the nation, is implementing “a moratorium on student-facing generative artificial intelligence (AI) in grades 2K–8, with limited use in high schools, and screen time restrictions depending on grade” for the 2026-27 school year.[1] The architecture is developmental staging — the restriction tightens as the student gets younger, on the theory that the earliest grades are where foundational skills are being built and a machine that does the thinking prevents the thinking from ever forming:

    • Grades 2K-2: student-facing generative AI banned, and no one-to-one screen time at all.
    • Grades 3-5: banned, with a recommended cap of 30 minutes of daily individual screen time.
    • Grades 6-8: banned, with a recommended cap of 45 minutes.
    • Grades 9-12: limited to approved, vetted programs, with screen time set school by school.[1]

    The carve-outs are as revealing as the ban. Across all grades, the policy does not apply to assistive technology provided through a student’s Individualized Education Plan or 504 Plan, and students with disabilities and English Language Learners keep access to the tools they need.[1] That distinction is the policy’s core logic in miniature: NYCPS is not restricting AI as a technology, it is restricting AI as a substitute for the student’s own cognitive work. The district describes generative AI as systems that “can produce inaccurate or biased information and can make it easier for students to bypass thinking they need to do themselves,” and explains that in younger grades the technology “can interfere with human interaction and hands-on learning that we know is developmentally appropriate for children.”[1] A screen reader that gives a dyslexic student access to text is a ramp; a chatbot that writes the essay is an elevator around the skill itself. The moratorium bans the elevator and keeps every ramp.

    High school is where the district’s other theory of the problem appears. All high school students must complete two 45-minute AI literacy modules covering how AI works, data privacy and bias, and appropriate uses of AI in school. Students in career-readiness pathways such as Career and Technical Education and FutureReadyNYC may use specific AI tools, but only with teacher supervision, and five centrally approved pilots — Quill, Edia, Brisk Teaching, Playlab, and Intel AI-Ready Schools — are the only avenue for classroom use, with each student limited to one pilot and dosage kept deliberately small: Quill, for example, is capped at roughly 15 minutes per week.[1] The selection criteria required each tool to pass the district’s data privacy review and to support students “with writing or math where the student does their own thinking.”[1] In other words, the district did not merely permit AI for older students; it curated a category of AI designed to preserve the cognitive work the ban exists to protect.

    Chancellor Kamar Samuels compressed the whole policy into a single sentence: “We are putting safety first and ensuring technology serves learning—not the other way around.”[1]


    The Revolt That Wrote the Rule

    The final policy reads as considered and deliberate. Nothing about how it came to be was. In March 2026, the city released its long-awaited initial AI guidance, built around green, yellow, and red light designations for different uses — and, in practice, largely leaving student use up to individual teachers.[2] The proposal was met with immediate backlash: hundreds of parents testified against it, city councilors signed a letter saying it did not go far enough, and teachers organized for a moratorium.[2] At a marathon May 1 Panel for Educational Policy meeting, scores of parents and educators demanded a full moratorium on AI use in city schools, and more than half the members of the City Council signed onto the push.[3]

    Samuels acknowledged in May that the city had “missed the mark” with its initial guidance and signaled the final version would carry more stringent rules, especially for younger students.[3] Officials originally promised the revision for June, then told the City Council it was delayed — with no updated timeline.[3] Then came the step that made the retreat concrete: in a Monday email to principals first reported on July 8, Samuels asked every school to hold off on purchasing educational software until the AI guidance was finalized. A department spokesperson framed it as protocol-building: “This necessary pause is part of our work to ensure proper protocols are in place while supporting our students and staff and making sure that any digital tool used in classrooms is properly assessed for safety and privacy.”[3]

    The purchasing freeze exposed the governance problem underneath the AI debate. At a June hearing, department officials struggled to answer basic questions about how many schools were using AI products and which products they were — because purchases are largely made at the school level and are not always tracked, an issue a state comptroller audit had already flagged as a significant student data privacy gap.[3] The district has since surveyed schools on what software they actually use.[3] The core logic of the freeze is therefore visibility, not ideology: a school system cannot set rules for tools it cannot inventory, and the moratorium is as much an act of procurement hygiene as of pedagogy.

    The freeze has a cost. Because schools must submit new purchase orders annually even for software they have used for years, the pause reached into budgets and intervention plans already built for the fall. One Manhattan principal, speaking anonymously: “We have already built budgets, planned interventions, and made decisions about programs for next year. Changing direction in July makes summer planning significantly more complicated and creates uncertainty about whether schools will have access to the tools they planned to use.”[3] That principal’s complaint is the moratorium’s design constraint stated plainly — a policy built mid-summer lands on schools that had already planned around the old rules.


    The National Pattern: Converging Caution, Diverging Instincts

    New York is the largest district to restrict student-facing AI, but it moved in a national current. Ahead of the 2026-27 school year, many districts adopted a cautionary posture — guardrails and AI education, but few specific system-wide rules, with in-classroom decisions delegated to individual teachers and at-home use largely untouched — while others approved no guidelines at all.[2] The vacuum is measurable: 60 percent of educators said their districts’ AI policies were unclear in a 2025 survey, and just 18 percent of teachers reported receiving any formal guidance from administrators on how AI tools should be used.[2] The opacity reaches students directly — seven in 10 teenagers use AI for schoolwork, but less than a third have discussed what AI is and how it works with a teacher in class, and only 44 percent have received any guidance on when it should or should not be used.[2]

    Facing the same technology, the country’s biggest districts reached for three different instruments:

    • Chicago trusts process. The nation’s third-largest district released a generative AI guidebook emphasizing AI literacy, a “co-pilot” framing, and an approved-tool list that pointedly excludes ChatGPT and Claude. But a parent group, Illinois Families for Public Schools, has now secured commitments from half of the candidates on the fall school board ballot to go much further: a three-year moratorium, screen time restrictions, and a chief privacy officer.[2]
    • Houston trusts preparation. The district partnered with OpenAI to give students and teachers ChatGPT Edu, and is converting nine campuses into “Future 2” schools with AI-focused programming — with up to 100 schools set to adopt the model by July 2031.[2]
    • New York, after its revolt, trusts restriction. Katy Independent School District outside Houston showed the same instinct in miniature, banning generative AI tools for K-6 students in July.[2]

    Los Angeles Unified supplies the cautionary tale that haunts all three approaches. The district released AI guidelines in 2024 and launched a proprietary chatbot named “Ed” — then shut it down three months later when its developer went out of business. In February, Superintendent Alberto Carvalho had his home raided by the FBI, reportedly in connection with the company behind the $3 million chatbot, and in June the district banned screen time before second grade.[2] The lesson districts absorbed is that AI procurement is not just an instructional decision but a liability decision, and that a vendor’s collapse can strand a half-million-student technology program overnight.

    Between the district experiments sits an industry of middlemen. The New York-based company AI for Education has helped more than 40 districts develop generative AI guidelines, and its CEO, Amanda Bickerstaff, argues the two halves of the problem are inseparable:

    “AI literacy cannot be done well without guardrails, and guardrails cannot be done well without AI literacy. Literacy must go hand in hand because the technology is too new. It works too differently than how you think it does, and it is too disruptive to be able to just come out with one or the other.”[2]

    Meanwhile the federal government is pushing in the opposite direction from New York’s moratorium. The April 2025 executive order “Advancing Artificial Intelligence Education for American Youth” allocates resources for AI literacy training, creates an AI-in-education task force, and directs the National Science Foundation to prioritize training initiatives that expand AI use — with Google, NVIDIA, and xAI among the companies committing resources to youth AI training under it.[2] The result is a policy sandwich: Washington accelerating adoption, the research establishment pulling toward restriction, and the actual rules being written — if they are written at all — one school board at a time.


    The Research Beneath the Ban

    The moratorium did not emerge from a pedagogical vacuum. The most consequential line in the field this year came from the Brookings Institution’s Center for Universal Education, which ran a yearlong global study — a deliberate “premortem” of AI in education before the failures can accumulate — built on interviews and consultations with over 500 students, teachers, parents, education leaders, and technologists across 50 countries, a review of over 400 studies, and a Delphi panel:[4]

    “At this point in its trajectory, the risks of utilizing generative AI in children’s education overshadow its benefits. This is largely because the risks of AI differ in nature from its benefits—that is, these risks undermine children’s foundational development—and may prevent the benefits from being realized.”[4]

    That asymmetry is the intellectual spine of New York’s grade line. If AI’s benefits are enhancements and its risks are damages to the developmental process itself, then the two do not cancel out — a benefit forgone can be recovered later, but a foundation not built may never be. Brookings distilled its findings into three pillars — prosper, prepare, protect — and twelve recommendations, including one that reads like a description of New York’s pilot criteria: “Use AI tools that teach, not tell.”[4] The district’s requirement that pilot tools support students “where the student does their own thinking” is that recommendation translated into procurement language.[1]

    The student-side data explains why the city paired its ban with mandatory literacy rather than stopping at the ban. Common Sense Media’s Youth AI Safety Institute surveyed a nationally representative sample of U.S. teens in August 2026 and found that 70 percent use AI for schoolwork, and among those users 63 percent get an answer from AI — whether they use it as-is (25 percent), rewrite it to sound like themselves (31 percent), or improve it with what they know (35 percent).[5] Yet only 27 percent of teens say a teacher has ever talked with them about what AI is and how it works.[5] Use is nearly universal; instruction is nearly absent. Common Sense Media CEO Jim Steyer put the tension in one line: “Students understand that AI threatens learning, but the reality is that it is too tempting and convenient for them to look the other way. The question is whether adults can provide them with the skills they need to use it responsibly.”[5]

    The same survey carries the strongest warning for the moratorium’s design. Among teens who use AI for schoolwork, 44 percent have had an AI tool blocked on a school network or device — and 59 percent of those students simply switched to a personal device afterward, where there is even less support.[5] A ban at the schoolhouse door does not remove the technology from a teenager’s pocket; it relocates the usage to the least supervised environment available. That finding is why New York’s structure — hard restriction for the young, mandatory literacy modules for the old — is more coherent than either a pure ban or a pure embrace: the restriction governs the years when unsupervised use does the most developmental damage, and the literacy requirement is an attempt to make the high schooler’s personal device less dangerous.

    Teens themselves, notably, do not describe a simple cheating story. Among those who use AI, 66 percent say it helps them understand schoolwork rather than just finish it faster, 38 percent say it leads them to generate fewer of their own ideas, and 39 percent say they feel they are missing out on learning when they use it to complete assignments. Asked which skills matter most for the future, they named working well with others (55 percent), fact-checking and critical thinking (50 percent), and reading and understanding complex information (48 percent) — human skills, chosen over AI skills, by the generation using AI the most.[5]


    What to Watch

    • The formal rule behind the family page. The moratorium is live on the district’s family-facing page, but as of the July purchasing freeze the department had no public timeline for finalizing the underlying guidance, and the freeze itself was framed as temporary. Whether the formal rule matches the page — and when principals can buy software again — is the first test of the policy’s durability.[1] [3]
    • The five pilots. New York’s high school exception is really an experiment: five vetted tools, dosage caps of minutes per week, one pilot per student, and a district evaluation plan in development. It is the first district-scale test of whether “AI that teaches, not tells” can be procured on purpose rather than by accident.[1] [4]
    • Chicago’s school board elections. Half of the candidates on the fall ballot have pledged to a three-year generative AI moratorium and a chief privacy officer. If they win, the nation’s third-largest district would follow its largest into restriction — and the moratorium would stop being a New York story and become a big-city template.[2]
    • The statehouse translation. Most state rules now mirror what districts built independently — guardrails, literacy, workforce framing — and some have already proven difficult for districts to implement. The open question is whether any legislature converts New York’s line-through-eighth-grade into statute, which would extend the rule to districts that never had a parent revolt of their own.[2]
    • The home-use gap. No district instrument governs AI use at home, and the data says that is where blocked students go. The next phase of this fight is likelier to be about homework than classwork — and it is the one arena where a school system’s authority runs out.[5] [2]

    The moratorium semester is best read not as a verdict on AI but as a verdict on adult infrastructure. Only 27 percent of teens have ever had a teacher explain what AI is; the nation’s largest district could not, until this summer, say which AI products its own schools were using.[5] [3] New York’s answer is to buy time — ban the machine for the youngest, teach about it to the oldest, vet everything in between — on the bet that restriction is temporary and literacy is permanent. The 2026-27 school year is the first test of whether that bet can survive contact with a million backpacks containing a million unrestricted phones.


    References

    [1] [PRIMARY] New York City Public Schools. “Guidance on Artificial Intelligence (AI) and Screen Time.” 2026-27 school year policy. schools.nyc.gov

    [2] [SECONDARY] Rodrigo, Chris Mills. “In the US, Caution Rules on AI Ahead of K-12 School Year.” Tech Policy Press. August 31, 2026. techpolicy.press

    [3] [SECONDARY] Elsen-Rooney, Michael. “Kamar Samuels asks NYC schools to pause software purchases until AI guidance is final.” Chalkbeat New York. July 8, 2026. chalkbeat.org

    [4] [SECONDARY] Burns, Mary, Rebecca Winthrop, Natasha Luther, Emma Venetis, and Rida Karim. “A new direction for students in an AI world: Prosper, prepare, protect.” Brookings Institution, Center for Universal Education. January 14, 2026. brookings.edu

    [5] [SECONDARY] Common Sense Media (nonprofit research and ratings organization). “Teens Who Use AI for Schoolwork Are Handing Off Critical Thinking, Common Sense Media Research Finds.” August 18, 2026. commonsensemedia.org


    AI automated and generated content. Human reviewed.