Tag: NIST AI RMF

  • 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 Six months later — with the promised request for information still unpublished — Reuters documented examiners at large and small banks alike asking a much blunter question: how do you turn your AI agents off?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 Examiner’s Playbook: How a Footnote Carve-Out, a Missing RFI, and a Three-Tier Worksheet Produced the First Supervisory Rules for AI Agents in American Banking

    The Examiner’s Playbook: How a Footnote Carve-Out, a Missing RFI, and a Three-Tier Worksheet Produced the First Supervisory Rules for AI Agents in American Banking

    On September 16, the Conference of State Bank Supervisors released a five-part framework for examining AI at banks and nonbank lenders — including the generative and agentic systems that federal banking agencies wrote out of their model risk guidance in April and have not yet replaced. It creates no new legal obligations. It also previews exactly what examiners will ask.


    What Was Released

    The CSBS Artificial Intelligence Supervisory Framework is a toolkit for state examiners, five documents designed to be used together: a Core Examiner Guide with scoping questions and document request lists; an Examiner Work Program that expands each procedure; Nonbank AI Supplements for lenders and other non-depository companies; an AI Use Case Risk Tiering Worksheet; and a Source Support Document identifying the materials the framework draws on.[1] The framework is explicitly discretionary — “a tool,” not a rule — and each state financial regulator decides how much of it, if any, to fold into its own exam program.[2]

    It was approved by CSBS’s State Supervisory Processes Committee and Non-Depository Supervisory Committee in August 2026 and built on existing public resources: the National Institute of Standards and Technology’s AI Risk Management Framework, the Cyber Risk Institute’s Financial Services AI Risk Management Framework, and the U.S. Treasury’s AI Lexicon.[2] CSBS President and CEO Brandon Milhorn framed the purpose as enabling adoption, not restraining it:

    “The use of AI provides a powerful new tool for financial institutions to improve services, protect consumers, and increase operating efficiency. While any new technology can present risks, the CSBS AI Supervisory Framework provides a principles-based approach to governance intended to help financial institutions explore and implement AI with additional confidence.”[1]

    That framing — governance as a condition of confidence — is the framework’s operating logic. A bank that can show an examiner a documented inventory, a tiering rationale, and controls matched to risk has a path to deploy aggressive AI; a bank that cannot is flying on tolerance alone.


    The Gap It Fills

    The framework exists because of a footnote. On April 17, 2026, the OCC, Federal Reserve, and FDIC issued revised interagency guidance on model risk management, replacing the 2011 guidance (SR 11-7) that had governed bank modeling for fifteen years, rescinding the OCC’s handbook booklet and the 2021 BSA/AML model risk statement along the way.[3] The revised guidance narrows the definition of “model” to complex quantitative methods underpinned by statistical, economic, or financial theories — excluding spreadsheet arithmetic and deterministic rule-based software — and then, in a single footnote, removes the newest technology of all: “Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance.”[3]

    Federal Reserve Vice Chair for Supervision Michelle Bowman confirmed the reading weeks later: “The revised guidance now applies narrowly to traditional models and basic AI applications. Going forward, we expect other risk-management and governance practices to support adoption of generative and agentic AI in ways that will encourage ongoing innovation.”[4] The agencies promised more to come — “the agencies plan 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” — but as of early October 2026, that RFI had not been published.[3][5]

    The result is a structural orphaning. The carve-out moved generative and agentic systems out of the model risk program without moving them out of the examination: a large language model drafting a suspicious activity report narrative, or an agent that gathers KYC documents and opens cases, still carries the financial, compliance, and conduct risk that model risk management exists to contain — but the guidance no longer supplies a method for governing them.[5] The American Bankers Association (an industry trade group) welcomed the carve-out as regulatory clarity and published member FAQs in May, while noting that the guidance itself directs banks to let their “risk management and governance practices” determine controls for anything not covered.[6] Clarity about scope still leaves each bank to define the controls.


    What Examiners Will Actually Ask

    The CSBS Core Examiner Guide walks an examiner from initial scoping questions through document requests to procedures covering AI governance and oversight, AI inventories and use cases, and generative AI and emerging uses.[2] The companion Examiner Work Program, a 28-page document, expands each procedure — and its treatment of agentic systems is the first operational supervisory guidance any U.S. banking regulator body has published. It directs examiners to evaluate an AI agent’s permission boundaries, human checkpoints, logging, reversibility, and capability to halt actions when needed.[7]

    The logic of that checklist is a shift in control surface. For a traditional model, the question is whether the output is right; for an agent, the question is what the system can touch. Permission boundaries, reversibility, and a halt capability are controls on action, not on prediction — the supervisory vocabulary moves from statistical validation toward something closer to access management.

    The framework’s model risk supplement makes the same move for generative systems. It instructs examiners to ask whether an institution has defined each model’s intended use and limitations, addressed drift and performance degradation, applied meaningful human testing and escalation, and independently validated vendor outputs — and it specifically flags confabulation, overreliance, and non-repeatable outputs as failure modes that traditional model risk frameworks were not built to catch.[7] A model that gives a different answer to the same question is not a minor validation wrinkle; it breaks the repeatable-input-output assumption that the entire 2011-era framework rests on.


    The Three-Tier Worksheet

    The framework’s most concrete artifact is the AI Use Case Risk Tiering Worksheet, an optional tool that assigns each AI use case to one of three tiers based on four factors: consumer impact, human oversight, harm potential from errors or outages, and data sensitivity.[8][7]

    • Tier 1 — Low risk. AI is limited to internal use, outputs are subject to human review, consumer impact and data sensitivity are limited, and potential harm from errors or outages is low. The premise is that a human in the loop absorbs most failure modes, so baseline controls suffice.
    • Tier 2 — Moderate risk. AI takes a consumer-facing or decision-support role, with exception-based human oversight and moderate data sensitivity and harm potential. Oversight here is triggered rather than continuous — humans review the cases the system flags, which concentrates attention where the system itself signals uncertainty.
    • Tier 3 — High risk. Use cases involve direct consumer outcomes, sensitive personal data, limited human review, significant operational reliance, or material potential harm from errors or outages. These are the cases where an AI failure reaches a customer’s money or file without a person in between.

    The suggested controls are cumulative: a Tier 3 use case assumes the lower tiers’ controls plus its own. At every tier, the baseline expectations include a documented risk tiering, an inventory of AI use cases naming business owners and purpose, a written acceptable-use policy, periodic governance review, vendor contracts that address AI risks and include audit rights, staff training on the tools’ limitations, data quality controls, and periodic reporting to management.[7] At the top, the controls escalate to independent model validation and AI-specific incident response procedures.[8]

    The core principle is that oversight scales with decision proximity and reversibility, not with the technology label. The same foundation model can be Tier 1 drafting internal memos and Tier 3 approving loans — the worksheet forces institutions to tier the use, not the model, which is exactly the question the federal carve-out left unanswered.


    The Nonbank Reach

    Three Nonbank AI Supplements extend the framework to the non-depository companies state banking departments license and supervise — mortgage companies, money services businesses, consumer finance companies, and debt collectors. The supplements apply AI-specific overlays to existing review areas: third-party and vendor risk, model risk, and consumer protection.[7]

    The vendor supplement matters beyond nonbanks, because it fills a second federal gap. The 2023 Interagency Guidance on Third-Party Relationships takes a broad, principles-based approach that never addresses AI-specific considerations — and the proposed replacement the agencies issued on September 11, 2026 likewise does not reach AI-specific vendor risk.[7] The CSBS supplement directs examiners to ask which vendor relationships involve AI capabilities, whether due diligence covered data provenance, explainability limits and vendor dependencies, whether contracts address AI-specific issues, and whether the institution monitors vendor-provided AI outputs with contingency plans for failure.[7] The consumer protection supplement, in turn, asks whether data inputs — including proxies and alternative data — could produce inaccurate, inconsistent, or unfair consumer outcomes, and whether meaningful human review applies where AI outputs influence consumer-facing decisions.[7]


    Why a Discretionary Document Bites

    Nothing in the framework binds anyone. Three facts nonetheless make it hard to ignore:

    • It reaches most U.S. banks. State regulators supervise approximately 3,355 — about 79 percent — of the 4,233 FDIC-insured banks and savings institutions in the United States as of June 30, 2026.[7] A supervisory tool for state examiners is, numerically, a supervisory tool for American banking.
    • Adoption will be uneven. Each state agency decides independently how far to incorporate the framework, and there is no comment period or compliance deadline — so a multistate institution can face examiners applying different depths of AI review across states, potentially within the same examination cycle.[7] The New York State Department of Financial Services has already established its own AI supervisory regime, so the framework layers onto at least one state that moved first.[7]
    • CSBS frameworks have a history of becoming law. CSBS works with state regulators and legislators on model laws; its Money Transmission Modernization Act has been adopted in full or in part by almost two-thirds of the states.[7] The American Association of Residential Mortgage Regulators is already working with the Mortgage Bankers Association on a survey of how mortgage companies deploy AI, and will decide separately whether to adopt the framework for the nonbank mortgage industry.[7]

    The through-line is that discretionary documents in this system function as drafts of future expectations. An examiner’s question is soft today; the same question in a model law is hard tomorrow.


    The Enforcement Backstop

    The framework’s softness is cushioned by a harder floor arriving November 2, 2026. On September 1, the OCC and FDIC published a final rule that defines “unsafe or unsound practice” in regulation for the first time: conduct that departs from generally accepted standards of prudent operation and has caused, or is likely to cause, material harm to the institution’s financial condition, with matters requiring attention issuable at a lower “could reasonably be expected” threshold.[9] The Federal Reserve did not join the rule.[5]

    For OCC- and FDIC-supervised banks, this closes the loop the carve-out opened. An ungoverned AI agent with write access to payment workflows can produce losses that clear the materiality bar — and now there is a defined supervisory hook for criticizing the absence of controls, even though no AI-specific guidance exists to name them.[5] The unchanged layer beneath all of this is violations of law: the Equal Credit Opportunity Act and Regulation B still require specific reasons for adverse action whatever drafted the notice, the bank still owns every suspicious activity report narrative, and unfair-or-deceptive-practices rules still apply to what a customer-facing assistant says and sends.[5] The absence of AI rules is not the absence of liability.


    What to Watch

    • The interagency RFI. Promised in April, still unpublished in October. What it asks about agentic AI — and whether it adopts anything resembling the CSBS tiers — will determine whether federal guidance converges with the state playbook or diverges from it.[3][5]
    • The state adoption map. Which state banking departments incorporate the framework, how deeply, and on what timeline. Divergence across states is the near-term risk; convergence is the medium-term bet, on the CSBS model-law precedent.[2][7]
    • November 2. The unsafe or unsound practice rule takes effect, giving examiners a formal mechanism short of enforcement actions.[9] The first citation involving an AI system will be the test of whether “generally accepted standards of prudent operation” can mean anything when the governing guidance is silent.
    • The first Tier 3 examination. The worksheet is optional until an examiner applies it. The first public supervisory action premised on these tiers — likely at a state, not a federal, agency — will convert a discretionary document into the de facto standard for AI agents in banking.[8]

    References

    [1] [PRIMARY] Conference of State Bank Supervisors. “CSBS Announces AI Supervisory Framework.” September 16, 2026. csbs.org

    [2] [PRIMARY] Conference of State Bank Supervisors. “The CSBS Artificial Intelligence Supervisory Framework.” September 16, 2026. csbs.org

    [3] [PRIMARY] Office of the Comptroller of the Currency. “OCC Bulletin 2026-13: Model Risk Management — Revised Guidance.” April 17, 2026. occ.gov

    [4] [PRIMARY] Bowman, Michelle W. “Artificial Intelligence in the Financial System.” Board of Governors of the Federal Reserve System. Delivered April 27, 2026. federalreserve.gov

    [5] [SECONDARY] “After SR 26-2: A Control Framework for the Generative and Agentic AI Your MRM Program No Longer Covers.” Lumenova AI. October 6, 2026. lumenova.ai

    [6] [ADVOCACY] American Bankers Association (industry trade group). “Artificial Intelligence” — banking topic page and “FAQs on the Interagency Revisions to the Model Risk Management Guidance,” May 12, 2026. aba.com

    [7] [SECONDARY] Adams, David, David Alexander, Hope Anderson, Glen R. Cuccinello, Rachel Rodman, and Pratin Vallabhaneni. “CSBS Releases Artificial Intelligence Supervisory Framework.” White & Case LLP, via JD Supra. October 1, 2026. jdsupra.com

    [8] [SECONDARY] Ennis, Dan. “State regulators float AI framework for banks, examiners.” Banking Dive. September 17, 2026. bankingdive.com

    [9] [PRIMARY] Office of the Comptroller of the Currency and Federal Deposit Insurance Corporation. “Unsafe or Unsound Practices, Matters Requiring Attention.” 91 FR 56004. September 1, 2026. federalregister.gov


    AI automated and generated content. Human reviewed.


    Listen to the Podcast: AI Policy for the Bears

    Episode 33: The Examiner's Playbook: How a Footnote Carve-Out, a Missing RFI, and a Three-Tier Worksheet Produced the First Supervisory Rules for AI Agents in American Banking — click to listen

    Two bears break down this story on AI Policy for the Bears, Episode 33.

  • Bank Regulators Draw the Lines on AI Governance in 2026

    Bank Regulators Draw the Lines on AI Governance in 2026

    Federal banking agencies narrowed model risk guidance while signaling that formal rules for generative and agentic AI are on the horizon — and states, the CFPB, and the mortgage GSEs are not waiting.


    The Shift: From “Is AI Covered?” to “Which AI Is Covered?”

    In April 2026, the OCC, Federal Reserve, and FDIC issued revised supervisory guidance on model risk management that did something surprising: it explicitly excluded generative and agentic AI from its scope. The agencies narrowed the definition of “model” to require that it be a “complex” quantitative method, and stated that generative and agentic AI models are “novel and rapidly evolving” and therefore outside the guidance. Instead, the guidance applies to traditional statistical models and basic AI applications.[1]

    But the exclusion is not a hands-off green light. The revised guidance notes that excluded AI tools are not exempt from risk management broadly — banking organizations should still apply broader governance and risk management practices to them. The agencies also announced plans to issue a request for information in the near future on model risk management as it relates to banks’ use of AI, signaling that formal regulatory guidance for generative AI in banking is forthcoming.[1]

    The reasoning here matters. Federal regulators are acknowledging that the 15-year-old SR 11-7 framework, designed for structured quantitative models, cannot stretch to cover systems whose outputs are probabilistic text, images, or autonomous agent actions. Rather than force-fit a square peg, they narrowed scope and flagged a separate process — but they did not leave the excluded space unregulated. They left it governed by general risk management principles pending dedicated guidance.


    Bowman’s Three Principles for AI Supervision

    Federal Reserve Vice Chair for Supervision Michelle Bowman laid out the Federal Reserve’s thinking in a May 1, 2026 speech at a Financial Stability Oversight Council roundtable. She identified three principles regulators should follow when supervising emerging technology like AI:[2]

    • Stay abreast of new developments and coordinate across government. The pace of AI capability change — Bowman cited Anthropic’s “Mythos” model, which identifies cyber vulnerabilities, as an example — means regulators cannot rely on static frameworks that age out before they are finalized.
    • Regular communication regarding unique risks of novel and broadly impactful innovation. Bowman emphasized that supervisors must understand specific bank use cases: Is the AI performing material tasks? Does it have broad employee access or limited? Does it directly affect consumers, as in credit determinations?
    • The regulatory and supervisory response must accommodate evolution. Regulators should regularly review their approach, communicate with industry, and incorporate feedback from banks, service providers, and experts. Supervisory guidance must not hinder implementation — particularly for smaller banks that need access to innovation despite fewer resources.

    Bowman also noted that the Federal Reserve has shifted its supervisory focus over the past year to identifying and remediating material financial risk — prioritizing matters that could lead to bank failure — and that rapid AI adoption reinforces the need for adaptable guidance. She described the third-party risk-management guidance as having been “vague in its scope and application” for too long and said it is being updated and simplified.[2]


    The OCC’s Dual-Edged Warning

    In its Spring 2026 Semiannual Risk Perspective, the OCC warned that AI is “significantly transforming” the cybersecurity threat landscape for banks. The report found that AI can facilitate fraud while lowering the barrier to entry for threat actors and increasing the “speed, scale, and sophistication” of cyberattacks against financial institutions.[3]

    Yet the OCC also recognized a defensive role: “increasingly advanced AI tools coming into the market to assist with cybersecurity functions,” and noted that “a sound understanding of the potential benefits and possible risks associated with these advanced tools can be important for cyber risk management.” The report expressed support for banks’ “measured approach” to integrating generative and agentic AI into core operational and customer service functions, while maintaining guardrails and ensuring human oversight remains embedded in workflows.[3]

    The OCC further indicated that banks “may consider expanding their use of [generative AI] and agentic AI for material financial decisions” — but cautioned that more advanced forms of AI present governance challenges including “lack of explainability, data privacy and data poisoning issues, cybersecurity threats, and validation challenges where industry approaches are evolving.” The report stressed that “appropriate governance and risk management are essential for risk mitigation.”[3]


    The Tail Risk Dimension: Banks’ Exposure to AI Investments

    Banking AI governance is not only about banks using AI internally — it is also about banks financing the AI industry. A February 2026 Federal Reserve Bank of Chicago analysis examined banks’ direct exposure to AI-adjacent investments and the potential tail risk those exposures create.[4]

    The findings are striking in scale:

    • Large bank C&I commitments to AI-adjacent industries reached approximately $450 billion as of late 2025, up from about $250 billion in 2015 — rising from 9% to 13% of total commitments.
    • Average outstanding exposure to AI-adjacent industries is about 9% of Tier 1 capital at large banks; committed exposure averages 25%.
    • AI-company-issued debt stands at approximately $1.2 trillion (per a JPMorgan estimate), more than double the large bank C&I commitments.
    • Bank lending to data centers totaled $14.9 billion in the year through Q3 2025.

    The Chicago Fed’s core concern is interconnectedness: “In a tail risk event, stress in one AI-adjacent industry can spill over to multiple interconnected AI-adjacent industries.” The report identified a specific chain: 26% of C&I commitments to software companies ($50 billion) are rated B and below, up from 14% in 2019. Many of these lower-rated firms are unprofitable and rely on investor capital injections to subsidize AI spending. If capital injections decrease and rates stay elevated, borrowers could struggle, reduced infrastructure spending would follow, and knock-on effects would hit data centers, energy companies, and semiconductor manufacturers.[4]


    The CFPB’s Consumer Protection Anchor

    While bank regulators focus on safety and soundness, the Consumer Financial Protection Bureau has anchored its AI oversight in existing consumer protection law. The CFPB issued guidance establishing that lenders using AI or complex models must provide specific and accurate reasons for adverse actions — credit denials or changes to credit conditions. There is no special exemption for artificial intelligence.[5]

    The CFPB’s logic is that the Equal Credit Opportunity Act already requires specific adverse action explanations, and that requirement does not weaken simply because a black-box model makes it difficult to identify reasons. Creditors cannot point to a broad bucket — if a credit line is reduced based on behavioral spending data, the explanation must detail the specific negative behaviors, not just “purchasing history.” Selecting the “closest” factor from a sample checklist is non-compliant if it does not reflect the actual reason for the decision.[5]

    The CFPB also opened a whistleblower channel specifically for tech workers who design, develop, or implement AI systems used in consumer finance — a recognition that the people closest to how these systems work are often the first to see problems.[5]


    States and the Mortgage GSEs: Filling the Federal Gap

    With federal banking regulators still developing dedicated AI guidance, other actors are not waiting. The most consequential state-level development is the Colorado Artificial Intelligence Act (SB 24-205), the first U.S. state law imposing affirmative obligations on developers and deployers of high-risk AI systems. Enacted May 17, 2024 and effective February 1, 2026, the law covers consequential decisions in financial or lending services — among other domains — and requires deployers to notify consumers before AI-assisted decisions are made, provide reasons for adverse decisions, and offer opportunities to correct data and request human review.[6]

    The Colorado Act’s core duty is a reasonable care standard: both developers and deployers must use reasonable care to protect consumers from any known or reasonably foreseeable risk of algorithmic discrimination. Enforcement is exclusive to the Colorado Attorney General, with no private right of action. A rebuttable presumption of reasonable care is available to deployers who implement a risk management policy aligned with the NIST AI Risk Management Framework — making the voluntary NIST framework practically binding for any Colorado deployer seeking a legal safe harbor.[6]

    In the mortgage sector, the government-sponsored enterprises have moved directly. On December 3, 2025, Freddie Mac issued Guide Bulletin 2025-16, which became effective March 3, 2026 and imposed highly prescriptive governance, audit, and security requirements for seller/servicers using AI/ML. Fannie Mae followed on April 8, 2026, issuing its own Governance Framework effective August 6, 2026, requiring written policies, annual reviews, trustworthy system characteristics, third-party compliance equal to internal standards, and disclosure on request.[7]

    The GSE framework matters because Fannie Mae and Freddie Mac act as bellwethers for the mortgage industry. Their directives do not just apply to the largest lenders — they reach any approved seller/servicer using AI/ML in loan origination or servicing, which means most of the U.S. mortgage market. As one analysis noted, “as AI/ML models grow more complex and more deeply embedded in critical processes, seller/servicers must ensure these technologies are deployed safely, legally, ethically, and in alignment with Fannie Mae’s expectations.”[7]


    The International Layer: FSB and NIST as Common Reference Points

    Bowman, in her capacity as chair of the Financial Stability Board’s Standing Committee on Supervisory and Regulatory Cooperation, indicated that the FSB will publish a consultation draft report on AI practices in Q3 2026. The report is expected to include balanced analysis of benefits and challenges, key principles, and examples of successful AI deployment, and will be published for stakeholder comment.[2]

    The FSB workstream is also considering consistency in AI use expectations internationally — particularly in cybersecurity and critical infrastructure, where international supervisory expectations may conflict with home country expectations. U.S. Treasury and SEC colleagues are working closely on this effort.[2]

    Underlying both domestic and international efforts is the NIST AI Risk Management Framework, released in January 2023 as a voluntary standard. The Colorado AI Act’s explicit citation of NIST AI RMF as a benchmark for reasonable care — and the rebuttable presumption it creates — shows how a voluntary framework can become effectively mandatory when referenced by law. This dynamic makes NIST AI RMF a de facto common reference point across federal guidance, state law, and international coordination.[6]


    What This Means for Financial Institutions

    The 2026 landscape presents a layered governance picture that financial institutions must navigate simultaneously:

    • Federal banking regulators have narrowed model risk guidance and signaled that dedicated AI guidance is coming via a forthcoming request for information. Institutions should not treat the exclusion of generative AI from SR 11-7 as permission to skip governance — the agencies explicitly expect broader risk management practices to apply.[1]
    • The CFPB has anchored its expectations in existing law: adverse action notices must be specific, AI is not exempt, and tech workers are a recognized whistleblower channel.[5]
    • State laws — with Colorado’s as the template — impose affirmative duties of reasonable care, impact assessments, consumer notification, and human review rights, enforced by state attorneys general.[6]
    • Mortgage GSEs have imposed prescriptive governance requirements that reach the entire seller/servicer network, with Freddie Mac’s already in effect and Fannie Mae’s effective August 2026.[7]
    • Systemic risk is being monitored not just through bank use of AI, but through bank financing of the AI industry itself — with the Chicago Fed quantifying exposures and identifying interconnectedness-driven tail risk scenarios.[4]

    The throughline across these developments is that no single regulator or framework covers the full picture. A bank using generative AI for customer service, deploying a model for credit underwriting, lending to AI companies, and selling loans to Fannie Mae now faces expectations from at least five distinct governance layers — each with different standards, different timelines, and different enforcement mechanisms. The institutions that proactively build AI governance frameworks now, aligned with NIST AI RMF as a common reference point, will be better positioned to satisfy all of them when the forthcoming federal guidance arrives.


    References

    [1] Sullivan & Cromwell LLP — “Federal Banking Agencies Issue Revised Guidance on Model Risk Management” (April 29, 2026) [SECONDARY]

    [2] Federal Reserve Board — “Speech by Vice Chair for Supervision Bowman on Artificial Intelligence in the Financial System” by Michelle W. Bowman (May 1, 2026) [PRIMARY]

    [3] Consumer Finance Insights — “OCC Report Signals AI Governance Guidance Is on the Horizon as Banks Navigate Dual-Edged Risks” (May 19, 2026) [SECONDARY]

    [4] Federal Reserve Bank of Chicago — “Tail Risk for Banks Posed by Investments in Generative Artificial Intelligence” by Greg Cohen, Cooper Killen, and Simon Lau (February 2026) [PRIMARY]

    [5] Consumer Financial Protection Bureau — “CFPB Issues Guidance on Credit Denials by Lenders Using Artificial Intelligence” [PRIMARY]

    [6] Skadden, Arps, Slate, Meagher & Flom LLP — “Colorado’s Landmark AI Act: What Companies Need To Know” (June 2024) [SECONDARY]

    [7] Alston & Bird Consumer Finance Blog — “Fannie Mae Issues Governance Framework on Use of Artificial Intelligence and Machine Learning” by Stephen Ornstein (June 5, 2026) [SECONDARY]


    AI automated and generated content. Human reviewed.

  • The Federal-State Collision Over AI Hiring Regulation

    The Federal-State Collision Over AI Hiring Regulation

    The Trump administration’s Executive Order 14365 aims to sweep away state AI hiring laws in the name of national dominance, even as states from California to New Jersey accelerate their own regulations and a landmark lawsuit against Workday tests whether AI vendors can be sued as employers.


    The Federal Retreat

    The federal government’s approach to AI in employment has reversed course entirely since January 2025. Under the Biden administration, the Equal Employment Opportunity Commission launched its Initiative on Artificial Intelligence and Algorithmic Fairness in 2021, issued technical guidance in 2023 applying Title VII principles to AI-assisted hiring, and filed amicus briefs supporting plaintiffs in emerging AI discrimination cases.[1] That infrastructure is now gone.

    On January 20, 2025, President Trump revoked Biden’s executive order on the safe and trustworthy development of AI. Three days later, Executive Order 14179 directed federal agencies to review and roll back existing AI policies and regulations.[2] The EEOC removed its 2023 guidance on responsible AI use in employment selection from its website, the Department of Labor’s OFCCP retracted AI and equal employment opportunity guidance for federal contractors, and the agency deprioritized enforcement of disparate impact theories of liability.[2][3]

    On April 23, 2025, Executive Order 14281, titled “Restoring Equality of Opportunity and Meritocracy,” went further, disavowing federal reliance on disparate impact liability “in all contexts to the maximum degree possible” and directing agencies to deprioritize enforcement of disparate impact statutes and regulations.[4] The Biden-era EEOC guidance for employers using AI no longer exists as an enforceable reference point. As Fisher Phillips noted, no federal AI workplace law is expected from the Republican-controlled Congress in 2025 or 2026.[5]


    EO 14365: The Preemption offensive

    The administration’s most aggressive move came on December 11, 2025, with Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence.” The order frames state-level AI regulation as the primary threat to American AI innovation, arguing that “a patchwork of 50 different regulatory regimes” makes compliance more challenging, particularly for startups.[6]

    The order’s language is unusually pointed for a presidential directive. It specifically calls out a Colorado law banning algorithmic discrimination, warning that such laws “may even force AI models to produce false results in order to avoid a ‘differential treatment or impact’ on protected groups.”[6] This is a striking claim: it treats bias mitigation as a form of compelled falsification rather than a civil rights safeguard. The underlying logic is that any regulation requiring AI outputs to account for demographic disparities amounts to government-mandated distortion of truthful results — a framing that puts anti-discrimination law and free-market AI development in direct opposition.

    EO 14365 establishes several enforcement mechanisms:

    • AI Litigation Task Force: Within 30 days, the Attorney General was directed to establish a task force to challenge state AI laws on grounds including unconstitutional regulation of interstate commerce and federal preemption. The task force was formally established on January 9, 2026, but no litigation has been initiated as of mid-2026.[7]
    • Commerce Department evaluation: Within 90 days, the Secretary of Commerce was to publish an evaluation identifying state laws requiring AI models to alter truthful outputs or compelling disclosures that violate the First Amendment. This evaluation was not publicly available by its March 2026 deadline.[7]
    • Funding conditions: States with “onerous” AI laws may be deemed ineligible for non-deployment funds under the BEAD broadband program, and federal agencies may condition discretionary grants on states agreeing not to enact or enforce conflicting AI laws.[6]
    • FCC and FTC proceedings: The FCC Chairman was directed to initiate a proceeding on a federal reporting and disclosure standard that would preempt conflicting state laws, and the FTC Chairman was to issue a policy statement on preemption of state laws mandating what the administration calls “deceptive conduct.” Neither action had been taken by mid-2026.[7]

    The pattern is notable: the executive order sets ambitious deadlines, but agency implementation has lagged. As the employment law firm Law and the Workplace observed, the order “does not immediately impact existing AI laws but establishes a process to discourage, challenge, and potentially preempt them.”[7] Employers remain obligated to comply with state and local AI laws in the meantime.


    States Fill the Void

    Even as the federal government retreats, states are accelerating. The National Conference of State Legislatures reports that all 50 states, Puerto Rico, the U.S. Virgin Islands, and Washington, D.C. introduced AI-related legislation in 2025, with approximately 100 measures adopted across 38 states.[8] Several of the most consequential state-level moves in the employment domain occurred in late 2025 and early 2026:

    • California: In October 2025, the Civil Rights Council enacted regulations making it unlawful to use automated decision systems or selection criteria that discriminate against applicants or employees — or a class of applicants or employees — based on protected characteristics under the Fair Employment and Housing Act.[1] Separately, California’s CCPA automated decision-making technology (ADMT) rules take effect January 1, 2027, requiring pre-use notice, opt-out rights, and access to information about how ADMT is used in significant employment decisions.[9]
    • New Jersey: In December 2025, the Division on Civil Rights adopted administrative code clarifying that employers may face liability for “algorithmic discrimination” even without discriminatory intent, and even when relying on third-party developers. The regulations provide examples of violations, including AI tools that use current employee populations as baselines for candidate searches or that screen candidates based on availability.[1]
    • Illinois: Effective January 1, 2026, the Human Rights Act was amended by statute to codify as a civil rights violation the use of AI in changing terms or conditions of employment if it subjects a person to discrimination based on protected characteristics, and it prohibits using ZIP codes as a proxy for protected characteristics.[1]
    • Connecticut: On February 25, 2026, the Attorney General’s Office issued a memo invoking existing state anti-discrimination laws as applicable to AI in hiring and employment. Pending Senate Bill 435 would impose disclosure requirements, bias audits by government-approved third parties, and mandatory human review of automated employment decisions.[1]
    • Colorado: The state’s AI Act, effective for employment obligations on January 1, 2027, requires businesses to provide descriptions of ADMT’s role in adverse decisions within 30 days, offer opportunities for human review, and allow requests for corrections to factually incorrect personal data.[9]

    The core logic driving these state efforts is that existing civil rights protections do not vanish simply because a decision is made by an algorithm rather than a human. As the joint statement from the EEOC, DOJ, CFPB, and FTC put it in April 2023, “There is no AI exemption to the laws on the books.”[5] State regulators are applying that principle: they are not creating new protected categories but clarifying that old ones — race, sex, age, disability — apply when the decision-maker is a machine learning model. The disagreement with the federal government is not over whether discrimination is wrong, but over whether algorithmic bias counts as discrimination in the first place.


    Mobley v. Workday: The Defining Test Case

    While legislators and executive branch officials argue over the regulatory framework, the courts are quietly building the case law that may matter most. Derek L. Mobley v. Workday, Inc., pending in the Northern District of California, is the highest-profile test of whether AI hiring tools can produce liability under existing anti-discrimination statutes.[10]

    The lead plaintiff, Derek Mobley, is an African American man over age 40 who alleges depression and anxiety as disabilities under the ADA Amendments Act. He claims that since at least 2017, he applied for over 100 positions through employers using Workday’s platform and was rejected for every application.[10] Three additional named plaintiffs have joined, each bringing claims based on their own rejections through Workday-powered systems.

    The case turns on a question that could reshape the entire AI hiring industry: can an AI vendor be treated as an “employer” under federal anti-discrimination law? Workday argued it was not an employer and therefore not subject to liability. The court rejected that argument, finding the complaint adequately alleged that Workday is an agent of its client-employers, falling within the definition of “employer” under Title VII, the ADEA, and the ADA.[10] On May 16, 2025, the court granted preliminary certification of a collective action on the ADEA claim.[11]

    The plaintiffs’ theory of how bias enters AI hiring systems is particularly important for policy. They argue that bias creeps in through training data composition, the selection and weighting of input variables, the design of model objectives, and the evaluation criteria used to assess candidate “fit.”[10] Even where protected characteristics are not explicitly provided to the system, they can be inferred through proxy variables: employment gaps may reflect disability, years of experience may stand in for age, educational background may correlate with race, and application data may encode gender or caregiving responsibilities.[10] This is the mechanism that makes algorithmic discrimination both pervasive and hard to detect — and it is the reason state laws like Illinois’s ban on ZIP code proxies and New Jersey’s prohibition on using current employee populations as baselines target specific technical practices rather than abstract harms.

    The political context has shifted under the case. Under the Biden administration, the EEOC filed amicus briefs supporting the plaintiffs. Under the Trump administration, that guidance was removed from the EEOC website and the agency deprioritized disparate impact cases.[10] But the case continues regardless — individual litigants are not restricted from pursuing disparate impact claims, even when the federal government declines to participate.[10]


    A New Legal Theory: FCRA and the Consumer Report Problem

    Beyond disparate impact, a new theory of liability for AI hiring tools emerged in January 2026. A nationally recognized plaintiffs’ employment class action firm filed a putative class action against an AI recruiting and “talent intelligence” technology company, alleging that the company assembles public information about candidates from across the web to create a proprietary database, then sells prospective employers reports to evaluate candidates — creating and selling “consumer reports” without complying with applicable laws.[1]

    The claims asserted include violations of the federal Fair Credit Reporting Act, California’s Investigative Consumer Reporting Agencies Act, and California’s Unfair Competition Law.[1] This represents an entirely new theory for challenging pre-employment AI tools: rather than arguing that the algorithm discriminates, the plaintiffs argue that the entire business model of scraping and selling candidate data triggers a federal disclosure and consent regime that most AI hiring companies have never considered. If successful, this approach would force AI hiring vendors into the same compliance framework that governs traditional background check companies — a potentially far-reaching expansion of liability that does not depend on disparate impact doctrine at all.


    What Comes Next

    The federal-state collision over AI hiring regulation is unlikely to resolve cleanly. Several dynamics will shape the next phase:

    • State enforcement will proceed regardless of federal posture. State attorneys general and civil rights agencies have their own enforcement authority, and states with private rights of action — like California, New Jersey, and Illinois — give individual plaintiffs the power to bring cases even when federal regulators sit on the sidelines.[1]
    • The preemption fight will move to the courts. EO 14365’s AI Litigation Task Force has not yet filed suits, but the legal arguments — dormant commerce clause, field preemption, conflict preemption — will take years to litigate. In the meantime, state laws remain in effect.[7]
    • Mobley v. Workday could establish vendor liability. If the court ultimately holds that AI hiring vendors are “employers” under Title VII, the ADA, and the ADEA, every company that sells AI screening tools faces potential exposure to the full range of employment discrimination claims.[10]
    • The FCRA theory expands the attack surface. By pulling AI hiring tools into the consumer reporting framework, plaintiffs’ attorneys have found a path that bypasses the disparate impact debate entirely — one that does not depend on proving bias, only on proving that the vendor failed to follow disclosure and consent requirements.[1]

    The underlying tension is structural. The federal government wants a single, minimally burdensome national standard to preserve American AI competitiveness. States want to protect their residents from algorithmic discrimination using tools that already exist in their legal codes. And employers are caught between two authorities pulling in opposite directions — one telling them to innovate freely, the other telling them to audit, disclose, and explain. The resolution will likely come not from legislation or executive orders, but from the slow accumulation of court decisions that define what AI vendors owe the people whose careers their algorithms shape.


    References

    [1] [SECONDARY] Signs, Christiana L. and Romeo, Samuel. “State AI Hiring Tool Regulations Filling Federal Void.” Reed Smith Employment Law Watch. April 22, 2026. https://www.reedsmith.com/our-insights/blogs/employment-law-watch/102mqfg/state-ai-hiring-tool-regulations-filling-federal-void/

    [2] [SECONDARY] Lada, Jennifer; Chapman, D’Andre; and Askinazi, Laura. “Artificial Intelligence in Hiring: Diverging Federal, State Perspectives on AI in Employment.” Holland & Knight Alert. March 20, 2025. https://www.hklaw.com/en/insights/publications/2025/03/artificial-intelligence-in-hiring-diverging-federal-state-perspectives

    [3] [SECONDARY] Meneghello, Richard R. and Walton, David J. “Comprehensive Review of AI Workplace Law and Litigation as We Enter 2025.” Fisher Phillips Insights. January 3, 2025. https://www.fisherphillips.com/en/insights/insights/comprehensive-review-of-ai-workplace-law-and-litigation-as-we-enter-2025

    [4] [PRIMARY] Trump, Donald J. “Executive Order 14281 — Restoring Equality of Opportunity and Meritocracy.” Federal Register. April 23, 2025. https://www.federalregister.gov/documents/2025/04/29/2025-07389/restoring-equality-of-opportunity-and-meritocracy

    [5] [SECONDARY] Meneghello, Richard R. and Walton, David J. “Comprehensive Review of AI Workplace Law and Litigation as We Enter 2025.” Fisher Phillips Insights. January 3, 2025. https://www.fisherphillips.com/en/insights/insights/comprehensive-review-of-ai-workplace-law-and-litigation-as-we-enter-2025

    [6] [PRIMARY] Trump, Donald J. “Executive Order 14365 — Ensuring a National Policy Framework for Artificial Intelligence.” Federal Register. December 11, 2025. https://www.federalregister.gov/documents/2025/12/16/2025-23092/ensuring-a-national-policy-framework-for-artificial-intelligence

    [7] [SECONDARY] “What President Trump’s AI Executive Order 14365 Means For Employers.” Law and the Workplace. April 2026. https://www.lawandtheworkplace.com/2026/04/what-president-trumps-ai-executive-order-14365-means-for-employers/

    [8] [PRIMARY] National Conference of State Legislatures. “Artificial Intelligence 2025 Legislation.” NCSL. July 10, 2025. https://www.ncsl.org/technology-and-communication/artificial-intelligence-2025-legislation

    [9] [SECONDARY] “AI Hiring Regulations: New State Laws, Risks & Compliance Guide (2026–2027).” SynHr. 2026. https://synhr.com/ai-hiring-regulations/

    [10] [SECONDARY] Edmondson, Elizabeth. “Emerging Liability for AI-Driven Hiring Tools: Key Developments in Mobley v. Workday, Inc.” Maynard Nexsen. April 23, 2026. https://www.maynardnexsen.com/publication-emerging-liability-for-ai-driven-hiring-tools-key-developments-in-mobley-v-workday-inc

    [11] [SECONDARY] “AI Bias Lawsuit Against Workday Reaches Next Stage as Court Grants Conditional Certification of ADEA Claim.” Law and the Workplace. June 2025. https://www.lawandtheworkplace.com/2025/06/ai-bias-lawsuit-against-workday-reaches-next-stage-as-court-grants-conditional-certification-of-adea-claim/


    AI automated and generated content. Human reviewed.

  • The 2026 State AI Hiring Law Wave: A Patchwork Compliance Guide

    The 2026 State AI Hiring Law Wave: A Patchwork Compliance Guide

    On January 1, 2026, three states launched divergent AI employment regimes — Illinois, Texas, and California — each betting on a different theory of how to govern algorithms that decide who gets hired. Together they reveal the core tension in AI hiring policy: transparency versus innovation, state authority versus federal preemption, and the difficulty of regulating bias you cannot see.


    Why 2026 Matters for AI in Hiring

    AI hiring tools processed more than 30 million job applications in 2024, and that volume triggered hundreds of discrimination complaints across federal and state agencies.[1] The problem is structural, not incidental: algorithms trained on historical hiring data learn the discriminatory patterns embedded in that data, and they keep discriminating even when protected characteristics like race and gender are stripped out, because the system finds proxy variables — graduation years stand in for age, ZIP codes for race, employment gaps for motherhood.[1] This is the “black box” problem: AI systems can be so complex that even their developers struggle to explain how a given decision was reached, which makes it nearly impossible for an employer to audit, defend, or even understand its own hiring outcomes.[1]

    Into that gap stepped the states. In the absence of comprehensive federal legislation, all 50 states plus Puerto Rico, the U.S. Virgin Islands, and Washington, D.C. introduced AI legislation in 2025, and roughly 100 measures were adopted across 38 states.[2] Three of the most consequential for employers took effect on January 1, 2026: Illinois HB 3773, Texas TRAIGA, and California’s automated-decision-system regulations. Each chose a different regulatory philosophy — and each leaves employers with a different compliance burden.


    Illinois HB 3773: The Broadest Net

    Signed by Governor J.B. Pritzker on August 9, 2024, and effective January 1, 2026, HB 3773 amends the Illinois Human Rights Act to regulate AI across the full employment lifecycle — recruitment, hiring, promotion, training selection, discipline, termination, and terms of employment.[3] The law defines AI sweepingly as “any machine-based system that, for explicit or implicit objectives, generates outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments,” with no carve-outs for specific AI types.[3]

    What makes Illinois distinctive is its effects-based prohibition: it is unlawful to use AI that has the effect of discriminating against protected classes, even if the discrimination is unintentional.[3] The law does not require formal bias audits or impact assessments — but that absence is a double-edged sword. Because the prohibition is effects-based, an employer accused of discrimination has no statutory safe harbor to point to. The practical implication is that proactive bias testing becomes essential for defense, even though it is not mandated.[3]

    The law also specifically bans the use of ZIP codes as a proxy for protected characteristics — a provision aimed squarely at the proxy-variable problem, where seemingly neutral geographic data encodes racial and socioeconomic information.[3][1] Employers must notify employees and applicants when AI is used for covered employment decisions, though the exact form and timing of that notice will be set by rules the Illinois Department of Human Rights is tasked with adopting.[3] Enforcement runs through the IDHR and the Illinois Human Rights Commission, with remedies including back pay, reinstatement, emotional distress damages, and attorneys’ fees.[3]


    Texas TRAIGA: The Minimalist Counterweight

    The Texas Responsible Artificial Intelligence Governance Act (TRAIGA, HB 149), effective January 1, 2026, takes nearly the opposite approach. It prohibits AI systems used for “restricted purposes” — including encouraging self-harm, infringing constitutional rights, unlawful discrimination, and generating child sexual abuse material — but it rejects disparate impact as a standalone basis for liability.[4] In practical terms, this means a Texas employer whose AI tool produces racially skewed outcomes is not automatically liable unless the discrimination was intentional, a far higher bar than Illinois’s effects-based standard.[4]

    TRAIGA grants enforcement authority exclusively to the Texas Attorney General, who must provide a 60-day cure period before pursuing penalties.[4] There is no private right of action — individuals cannot sue. Penalties range from $10,000 to $12,000 for curable violations up to $80,000 to $200,000 for incurable ones, with daily continuing-violation fines of up to $40,000.[4] The law offers an affirmative defense for organizations that comply with the NIST AI Risk Management Framework, effectively turning a voluntary federal standard into a liability shield.[4] Crucially, developers and deployers are not liable solely because end users misuse their systems — liability depends on intent in creating or distributing the AI.[5]

    The core logic of TRAIGA is a bet that over-regulation chills innovation more than under-regulation harms workers. By requiring intent, excluding private lawsuits, and building in a cure period, Texas has constructed a regime that is easy to comply with on paper but difficult to enforce in practice — particularly for the diffuse, hard-to-detect bias that proxy variables produce.


    California CRD Regulations: The Most Detailed Rulebook

    California’s approach occupies a middle ground but produces the most detailed compliance obligations of the three. On June 27, 2025, the California Office of Administrative Law approved final regulations from the Civil Rights Council clarifying how the state’s existing antidiscrimination laws apply to AI in employment — effective October 1, 2025.[6] These are regulations, not new statutes: they do not create new protected classes or new causes of action, but they specify in granular detail how the Fair Employment and Housing Act applies to automated-decision systems (ADS).

    The regulations establish that using an ADS may violate California law if it harms applicants or employees based on protected characteristics including gender, race, and disability.[6] They require employers to maintain employment records, including automated-decision data, for a minimum of four years.[6] They affirm that ADS assessments — including tests, questions, or puzzle games that elicit disability information — may constitute an unlawful medical inquiry under the ADA.[6] And they add definitions for “automated-decision system,” “agent,” and “proxy” to the regulatory lexicon.[6]

    As CRD Director Kevin Kish stated when the regulations were approved: “These new regulations on artificial intelligence in the workplace aim to help our state’s antidiscrimination protections keep pace.”[6] The regulations also clarify that vendors can be held liable under agency principles when they exercise control over employment decisions — a provision that extends legal exposure beyond the employer to the HR technology company itself.[1] This vendor-liability principle represents a structural shift: it means the firm that builds the algorithm shares legal responsibility for the outcomes it produces, not just the company that buys it.


    Colorado: The Comprehensive Standard Arriving June 30

    Colorado’s SB 24-205, the Colorado AI Act — the first comprehensive U.S. state law targeting high-risk AI systems — was originally set to take effect February 1, 2026, but was delayed to June 30, 2026.[5][7] It applies to entities doing business in Colorado that use high-risk AI for employment decisions affecting Colorado residents, regardless of where the employer is located.[7] It is the only state law specifically named in President Trump’s December 2025 executive order on AI preemption, which directed the Attorney General to challenge state AI laws deemed inconsistent with federal deregulatory policy.[5]

    Colorado requires reasonable care to prevent algorithmic discrimination, annual impact assessments, transparency notices, documentation of adverse decisions, and an appeal process.[7] Violations are treated as unfair trade practices under the Consumer Protection Act, with penalties up to $20,000 per violation.[4] There is no private right of action, but complainants must notify the Attorney General within 90 days before penalties can issue.[1] A narrow exemption applies to employers with fewer than 50 employees who do not use their own data to train or improve their AI systems.[7]


    The Federal Preemption Shadow

    Hanging over all of these state laws is the Trump administration’s deregulatory posture and explicit push for federal preemption. On January 20, 2025, President Trump revoked Biden’s Executive Order 14110, which had required safety-testing sharing, agency-level AI actions, Chief AI Officer appointments, and Department of Labor employment AI reviews.[4] A subsequent January 23, 2025 executive order declared U.S. policy to “sustain and enhance America’s global AI dominance.”[4]

    Then on December 11, 2025, Trump signed “Ensuring a National Policy Framework for Artificial Intelligence,” which proposes to preempt state laws deemed inconsistent with federal policy.[5] It directs the Attorney General to establish an AI Litigation Task Force to challenge state AI laws on commerce and preemption grounds, and tasks the Commerce Department — with a March 11, 2026 deadline — with identifying burdensome state laws, particularly those requiring altered truthful outputs or compelling disclosures that might violate the First Amendment.[5][4]

    More recently, on June 2, 2026, Trump signed Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security,” framing the federal posture as “innovation over regulation.”[8] Two days later, Representatives Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.) released a 269-page bipartisan discussion draft — the Great American AI Act of 2026 — that would preempt state laws regulating AI model development for three years while preserving state laws of general applicability such as privacy and anti-discrimination statutes.[4] Whether preemption passes Congress or survives judicial review remains uncertain. Until courts rule, the state laws remain enforceable — and the Baker Botts analysis cautions that organizations should continue complying with applicable state requirements while monitoring legal challenges.[4]


    Compliance Strategies for a Patchwork World

    For employers operating across state lines, the core challenge is that a single AI hiring tool may be subject to three or four different legal regimes simultaneously. The K&L Gates analysis recommends a set of practical steps that reduce risk regardless of which state’s law applies:[7]

    • Map your AI footprint. Identify every AI system currently in use or planned for hiring, promotion, or termination decisions. You cannot comply with a notice requirement you do not know about. This inventory should be living documentation, updated whenever a new tool is deployed or an existing one is substantially modified.[7]
    • Conduct bias testing — even where not required. Illinois does not mandate audits, but its effects-based prohibition makes bias testing essential for defense. Colorado and California effectively require it. Regular disparate-impact testing across protected classes creates a record of reasonable care that serves as both a compliance tool and a litigation shield.[3][7]
    • Ensure meaningful human review. California’s regulations require that human oversight be genuine — a trained reviewer empowered to override the AI, not a rubber stamp. NYC’s Local Law 144 already established that human review of AI rankings does not exempt the tool from audit requirements. The principle across jurisdictions is that a human who always agrees with the machine is not oversight.[1]
    • Document vendor compliance. California holds vendors liable under agency principles, and courts have ordered AI vendors to produce lists of all customers using discriminatory features — exposing entire customer bases to scrutiny. Employers must confirm that third-party AI providers meet legal and ethical standards, and maintain records of that confirmation.[7][1]
    • Retain records for four years minimum. California sets a four-year retention floor for automated-decision data. Even in states without explicit retention requirements, maintaining records of risk assessments, bias audits, mitigation steps, and vendor compliance creates the documentation that distinguishes a defensible process from an indefensible one.[6]

    The Structural Problem: Regulating What You Cannot See

    Beneath the technical compliance steps lies a deeper problem that none of these laws fully resolves. AI bias is often invisible by design: the system identifies correlations in training data that humans would never think to look for, and it acts on them. Removing protected characteristics from the input does not solve the problem because the algorithm finds proxies.[1] Graduation year becomes age. ZIP code becomes race. Employment gap becomes gender. The bias is real, but it is encoded in variables that appear neutral — which means an employer reviewing its own process in good faith may genuinely not know discrimination is occurring.

    This is why the vendor-liability development in California is significant. If the firm that builds the algorithm bears legal responsibility for its outputs, it creates an incentive to design for fairness at the point of creation rather than relying on downstream employers to detect problems they are not equipped to see. But it also raises a question that no state law yet answers clearly: when a vendor’s AI is trained on a customer’s historical data, who is responsible for the bias that data contains — the vendor who built the tool, or the employer who supplied the data?[1]

    The 2026 state law wave does not resolve these questions. It does establish, for the first time at scale, that employers cannot treat AI hiring tools as black boxes exempt from the discrimination law that governs every other employment decision. The compliance burden is real, the patchwork is messy, and the preemption fight is unresolved. But the baseline expectation has shifted: if you use AI to decide who gets hired, you are responsible for understanding what it does — and for proving that what it does is lawful.


    References

    [6] [PRIMARY] California Civil Rights Department. “Civil Rights Council Secures Approval for Regulations to Protect Against Employment Discrimination Related to Artificial Intelligence.” June 30, 2025. calcivilrights.ca.gov

    [3] [SECONDARY] Francis, Simone R.D., Benjamin W. Perry, and Lauren N. Watson. “Illinois Steps Up AI Regulation in Employment: Key Takeaways for Employers.” Ogletree Deakins. August 29, 2024. ogletree.com

    [5] [SECONDARY] King & Spalding. “New State AI Laws are Effective on January 1, 2026, But a New Executive Order Signals Disruption.” December 29, 2025. kslaw.com

    [4] [SECONDARY] Baker Botts. “U.S. Artificial Intelligence Law Update: Navigating the Evolving State and Federal Regulatory Landscape.” January 2026. bakerbotts.com

    [7] [SECONDARY] Rigney, Erinn L., et al. “Navigating the AI Employment Landscape in 2026: Considerations and Best Practices for Employers.” K&L Gates. February 3, 2026. klgates.com

    [1] [SECONDARY] “AI in Hiring: Emerging Legal Developments and Compliance Guidance for 2026.” HR Defense Blog. November 2025. hrdefenseblog.com

    [2] [SECONDARY] National Conference of State Legislatures. “Artificial Intelligence 2025 Legislation.” July 10, 2025. ncsl.org

    [8] [PRIMARY] The White House. “Promoting Advanced Artificial Intelligence Innovation and Security.” Executive Order 14409. June 2, 2026. whitehouse.gov


    AI automated and generated content. Human reviewed.

  • The Trump AI Executive Order: Removing Barriers to American Leadership

    The Trump AI Executive Order: “Removing Barriers to American Leadership”

    Signed January 23, 2025, this executive order rescinds Biden’s EO 14110 and pivots U.S. federal AI policy from a risk-management posture to an innovation-first, deregulatory approach.

    What Changed

    President Trump’s executive order, titled “Removing Barriers to American Leadership in Artificial Intelligence,” represents a sharp reversal from the Biden administration’s EO 14110 (October 2023). Key shifts:

    What Was Rescinded

    Biden’s EO 14110 had established:

    • Extensive safety testing requirements for frontier AI models (above a compute threshold)
    • Red-team testing and reporting obligations for dual-use foundation models
    • Federal agency AI governance requirements (Chief AI Officers, AI use case inventories)
    • Immigration provisions to attract AI talent
    • Standards for AI in federal procurement

    The Trump EO rescinded these requirements, calling them “Barriers to American Leadership.”

    What Was Kept or Added

    • Support for the U.S. AI Safety Institute (USAISI) — maintained, though with a revised mission focused on “promoting AI innovation”
    • A directive to develop an AI Action Plan within 180 days (by July 2025)
    • Review of all federal agency actions taken under Biden’s EO to identify and remove “undue burdens”
    • Commitment to maintain U.S. dominance in AI development

    The Policy Shift

    The contrast between the two administrations is stark:

    Dimension Biden EO 14110 Trump EO (Jan 2025)
    Posture Risk management, safety-first Innovation-first, deregulatory
    Frontier models Mandatory safety testing, reporting No mandatory testing
    Agency governance Chief AI Officers, inventories Review and remove “undue burdens”
    International engagement Aligned with Bletchley Declaration America-first framing

    Implications for AI Policy

    The Trump EO creates significant uncertainty for organizations that had begun implementing compliance programs under Biden’s EO:

    • Companies that established red-team testing and safety reporting protocols may no longer be required to use them
    • The NIST AI RMF remains voluntary but may receive less federal promotion
    • State-level laws like Colorado’s AI Act and NYC Local Law 144 become more important as federal requirements recede
    • The EU AI Act (Regulation 2024/1689) now stands as the world’s most comprehensive binding AI regulation, potentially creating a “Brussels Effect” where global companies default to EU compliance standards

    Primary Sources


    AI automated and generated content. Human reviewed. All policy text is summarized from the primary sources linked above.

  • The Rome Call for AI Ethics: 2024 Update — World Religions Sign On

    The Rome Call for AI Ethics: 2024 Update — World Religions Sign On

    The Rome Call for AI Ethics, originally promulgated in 2020 by the Vatican’s Pontifical Academy for Life, expanded significantly in 2024 with the signing by eleven world religions in Hiroshima, Japan. This updated article covers the original framework and the 2024 developments documented in the RenAIssance Foundation’s 2024 Mission Report.

    Rome Call 2024 Mission Report cover
    The RenAIssance Foundation 2024 Mission Report cover. View the full PDF report here.

    Origins (2020)

    The Rome Call for AI Ethics was promulgated on February 28, 2020, by the Pontifical Academy for Life, with the goal of creating a multi-stakeholder commitment to ethical AI development. The original document was signed by:

    • Pontifical Academy for Life
    • Microsoft
    • IBM
    • Fujitsu
    • Food and Agriculture Organization (FAO) of the UN
    • Italian Ministry for Innovation

    The Six Principles

    The Rome Call articulates six fundamental principles that should govern AI development and deployment:

    1. Transparency: AI systems must be explainable and their decision-making processes accessible to human understanding
    2. Inclusion: AI must serve all of humanity, with particular attention to marginalized and vulnerable populations
    3. Responsibility: Those who design and deploy AI systems bear responsibility for their impacts on individuals and society
    4. Impartiality: AI systems must not discriminate or perpetuate bias, and must treat all persons with equal respect
    5. Reliability: AI systems must function dependably and not cause harm through malfunction or error
    6. Security and Privacy: AI systems must respect user privacy and operate with appropriate data security safeguards

    Theological Foundation

    What distinguishes the Rome Call from secular AI ethics frameworks (like the NIST AI RMF or the EU’s ethics guidelines) is its explicit grounding in human dignity and the common good — core principles of Catholic social teaching. The document states that AI must “serve the human person and the human community” and that human beings must remain “the subject of AI, not the object.”

    2024: A Landmark Year — World Religions Sign the Rome Call

    The most significant development in 2024 was the interreligious signing of the Rome Call in Hiroshima, Japan, on July 9th–10th. The event, titled “AI Ethics for Peace,” brought together sixteen new signatories representing eleven world religions, including:

    • Buddhism (World Fellowship of Buddhists, Tendai Buddhism)
    • Hinduism (Dev Sanskriti University)
    • Sikhism (Guru Nanak Nishkam Sewak Jatha)
    • Confucianism (Sungkyunkwan)
    • Bahá’í International Community
    • Shinto traditions (Kurozumikyo, Konkokyo, Zenrinkyo)
    • Rissho Kosei-kai
    • Islamic, Jewish, and Christian leaders

    The event was greeted by Pope Francis with a dedicated message: “Recognizing the contribution of the cultural riches of peoples and religions in the regulation of artificial intelligence is key to the success of your commitment to the wise management of technological innovation.”

    The organizing partners — Pontifical Academy for Life, Religions for Peace Japan, Abu Dhabi Forum for Peace, and the Commission for Interreligious Relations of the Chief Rabbinate of Israel — also presented the Hiroshima Appeal, which urged the international community to use peaceful means to resolve all armed conflicts and reaffirmed that AI must be used only for the good of humanity and the planet.

    Other 2024 Milestones

    According to the RenAIssance Foundation’s 2024 Mission Report, the year saw several additional milestones:

    • January 1st — Pope Francis dedicated the 57th World Day of Peace to “Artificial Intelligence and Peace”
    • February 2nd — The Catholic Bishops’ Conference of India signed the Rome Call in Kerala, India
    • April 24th — Cisco (CEO Chuck Robbins) signed the Rome Call at the Vatican, becoming a major corporate signatory
    • April 30th — The Church of England signed the Rome Call
    • June 14th — Pope Francis mentioned the Rome Call and its principles at the G7 Session on Artificial Intelligence in Borgo Egnazia, Italy, calling for “algor-ethics” — ethical moderation of algorithms
    • October 29th — Seven universities of Region Campania, Italy, signed the Rome Call
    • November 12th — ConfProfessioni, representing self-employed entrepreneurs in Italy, signed the Rome Call
    • December 2nd — Lutech Group signed the Rome Call

    Pope Francis at the G7

    In his historic address to the G7 on June 14, 2024, Pope Francis articulated why the Rome Call matters:

    “An ethical decision is one that takes into account not only an action’s outcomes but also the values at stake and the duties that derive from those values. That is why I welcomed both the 2020 signing in Rome of the Rome Call for AI Ethics, and its support for that type of ethical moderation of algorithms and artificial intelligence programs that I call ‘algor-ethics.’”

    He emphasized that in a pluralistic global context, the Rome Call provides “a series of principles condensed into a global and pluralistic platform that is capable of finding support from cultures, religions, international organizations and major corporations.”

    The RenAIssance Foundation

    Established on April 12, 2021, by Pope Francis, the RenAIssance Foundation is the non-profit organization that promotes the Rome Call. Its vision: “technology will be at the service of human beings, respecting the dignity of each individual without causing any harm to humanity, protecting the common home that hosts every form of life.” The Foundation operates a Study Centre and has organized dissemination initiatives across cities including Washington, London, Paris, Madrid, Abu Dhabi, Hiroshima, Tokyo, Seoul, and New Delhi.

    Why It Matters for AI Policy

    The Rome Call represents a unique intersection of theology and technology policy. While not legally binding, it has influenced corporate AI ethics commitments (Microsoft, IBM, and now Cisco reference it in their AI principles) and provides a moral vocabulary for AI ethics discussions grounded in religious tradition. The 2024 expansion to include eleven world religions transforms it from a primarily Catholic initiative into a genuinely interreligious framework — arguably the most broadly endorsed faith-based AI ethics document in the world.

    For the theological and moral stratum of AI policy analysis, the Rome Call is a foundational document, and the 2024 Hiroshima signing marks its emergence as a global multi-faith standard.

    Primary Sources


    AI automated and generated content. Human reviewed. All policy text is summarized from the primary sources linked above. The cover image is from the RenAIssance Foundation 2024 Mission Report.

  • The Bletchley Declaration: 28 Nations Agree on AI Safety Priorities

    The Bletchley Declaration: 28 Nations Agree on AI Safety Priorities

    Signed at the AI Safety Summit on November 2, 2023, at Bletchley Park (the historic site of WWII codebreaking), the Bletchley Declaration established the first international consensus on AI safety risks and the need for collaborative governance.

    What It Says

    The Bletchley Declaration was signed by representatives of 28 countries including the United States, United Kingdom, China, the European Union, India, and others. It represents the first time so many nations formally acknowledged the need for international cooperation on AI safety.

    Key commitments in the Declaration include:

    1. Identifying AI Safety Risks of Shared Concern

    The Declaration identifies “frontier AI” (highly capable general-purpose AI models) as a category requiring special attention, noting risks including unintended consequences, loss of control, and potential misuse for harm at scale.

    2. Building a Shared Scientific Evidence Base

    Signatories committed to supporting “an internationally inclusive network of scientific research on frontier AI safety” — leading to the establishment of AI Safety Institutes in the UK, US, and other countries.

    3. Developing Risk-Based Policies

    The Declaration acknowledges that different countries may adopt different approaches but emphasizes the need for policies that are “interoperable” across jurisdictions — a principle also embedded in the EU AI Act‘s risk-based framework.

    What Followed

    The Bletchley Declaration spawned a series of follow-on summits:

    • Seoul AI Summit (May 2024): 16 companies committed to voluntary AI safety frameworks
    • UK AI Safety Institute: Established November 2023, the first government-backed institute dedicated to evaluating advanced AI systems
    • US AI Safety Institute (USAISI): Established through the NIST, works with the UK institute on shared evaluation methodologies

    Significance

    The Bletchley Declaration is significant for several reasons:

    • China’s participation: Despite geopolitical tensions, China signed the Declaration — a rare moment of US-China-EU alignment on technology policy
    • Focus on “frontier AI”: The Declaration was the first major international statement to specifically call out frontier models as a distinct risk category
    • Institutional legacy: It directly led to the creation of national AI Safety Institutes, which are now the primary government bodies evaluating advanced AI systems

    How It Relates to Other Frameworks

    The Bletchley Declaration is a diplomatic statement, not a binding regulation. It complements:

    • The EU AI Act — binding regulation with risk tiers
    • The NIST AI RMF — voluntary U.S. framework
    • The Trump administration’s AI EO (January 2025) — which rescinded Biden’s EO 14110 but maintained support for AI Safety Institute work

    Primary Sources


    AI automated and generated content. Human reviewed. All policy text is summarized from the primary sources linked above.

  • NIST AI Risk Management Framework: The Voluntary Standard Shaping US AI Policy

    NIST AI Risk Management Framework: The Voluntary Standard Shaping US AI Policy

    Published January 2023, the NIST AI Risk Management Framework (AI RMF 1.0) is the U.S. government’s flagship voluntary framework for managing AI risks. It has become the de facto reference standard cited in state laws, federal guidance, and private-sector AI governance programs.

    What It Is

    The NIST AI RMF is a voluntary, rights-preserving, non-sector-specific framework intended to help organizations manage risks associated with AI throughout the system lifecycle. It was developed through a multi-stakeholder process involving industry, academia, civil society, and government over two years.

    Core Functions

    The framework organizes AI risk management into four core functions, modeled on the NIST Cybersecurity Framework:

    1. Govern

    Policies, procedures, and organizational structures for AI risk management. This includes establishing roles, responsibilities, accountability structures, and organizational risk tolerance.

    2. Map

    Contextualizing AI systems and their risks — identifying the system’s purpose, intended use, stakeholders, potential benefits, and harms across the lifecycle.

    3. Measure

    Assessing, quantifying, and monitoring AI risks using appropriate metrics and methods. This includes evaluating trustworthiness characteristics like validity, reliability, safety, security, accountability, transparency, explainability, privacy, and fairness.

    4. Manage

    Prioritizing and acting on identified risks — allocating resources to mitigate the highest-priority risks, establishing response plans, and monitoring residual risk.

    AI RMF Playbook

    NIST also published a companion AI RMF Playbook providing practical guidance and suggested actions for implementing each subcategory of the framework. The Playbook is a living document, updated with community contributions.

    How It’s Being Used

    The NIST AI RMF has become the reference standard for U.S. AI governance:

    • Colorado AI Act (SB 24-205): Explicitly names the NIST AI RMF as an appropriate basis for a “risk management plan” — see SB 24-205 text
    • Federal agencies: Multiple agencies reference the AI RMF in procurement and internal AI guidance
    • Private sector: Companies including Microsoft, Amazon, and IBM have adopted the framework for internal AI governance

    Generative AI Profile

    In July 2024, NIST released a Generative AI Profile (NIST AI 600-1) as a companion resource specifically addressing risks unique to generative AI, including:

    • CBRN (chemical, biological, radiological, nuclear) information generation
    • Confabulation and hallucination
    • Dangerous, violent, or hateful content
    • Data privacy and confidentiality risks
    • Environmental impacts of large model training

    Primary Sources


    AI automated and generated content. Human reviewed. All policy text is summarized from the primary sources linked above.

  • FDA Framework for AI/ML-Enabled Medical Devices: Predetermined Change Control Plans

    FDA Framework for AI/ML-Enabled Medical Devices: Predetermined Change Control Plans

    The FDA’s approach to regulating AI/ML-enabled medical devices centers on a “total product lifecycle” framework and the concept of Predetermined Change Control Plans (PCCPs), allowing approved devices to evolve after market authorization without requiring new submissions for every update.

    The Framework

    The U.S. Food and Drug Administration regulates AI/ML-enabled software that meets the definition of a medical device under the Federal Food, Drug, and Cosmetic Act. As of 2025, the FDA has cleared or authorized more than 950 AI/ML-enabled medical devices through the 510(k), De Novo, and PMA pathways.

    The agency’s regulatory framework is built around several key guidance documents and concepts:

    Predetermined Change Control Plans (PCCPs)

    A PCCP allows a manufacturer to describe in their initial submission what types of modifications they plan to make to the AI/ML algorithm after deployment — such as retraining with new data or adjusting model parameters. If the FDA approves the PCCP, the manufacturer can implement those specific changes without filing a new submission for each update.

    This is a significant departure from the traditional medical device model, where any modification requires a new 510(k) or supplement. The FDA recognized that AI/ML algorithms are inherently iterative — they learn and adapt — and a traditional regulatory framework would stifle clinically beneficial improvements.

    Good Machine Learning Practice (GMLP)

    The FDA, in collaboration with Health Canada and the UK’s MHRA, published a set of Good Machine Learning Practice guiding principles in October 2021. These 10 principles cover:

    • Applying software engineering and quality system practices throughout the product lifecycle
    • Clinically robust training and testing datasets that reflect the intended patient population
    • Independent review of training data by qualified experts
    • Device performance monitoring and iterative improvement

    Transparency and Bias

    The FDA has emphasized the importance of transparency in AI/ML devices, particularly regarding:

    • How the device’s inputs and outputs should be interpreted
    • Limitations and contraindications
    • Populations for which the device has been tested and validated
    • Potential for bias if the training data does not represent the intended use population

    Relationship to EU AI Act

    Under the EU AI Act, most AI-enabled medical devices are classified as “high-risk” AI systems and are subject to conformity assessment requirements before being placed on the EU market. This includes risk management systems, data governance, technical documentation, and human oversight requirements. The FDA framework, by contrast, operates through device-specific clearance pathways rather than a horizontal AI-specific regulation.

    Primary Sources


    AI automated and generated content. Human reviewed. All policy text is summarized from the primary sources linked above.

  • Colorado Becomes First State to Pass Comprehensive AI Anti-Discrimination Law

    Colorado Becomes First State to Pass Comprehensive AI Anti-Discrimination Law

    SB 24-205, signed by Governor Jared Polis in May 2024, takes effect February 1, 2026. It is the first U.S. state law to impose affirmative obligations on developers and deployers of “high-risk” artificial intelligence systems.

    What the Law Requires

    Colorado’s Artificial Intelligence Act (Senate Bill 24-205) establishes a duty of care for anyone who develops or deploys a “high-risk AI system” — defined as any AI system that makes, or is a substantial factor in making, a “consequential decision” affecting a consumer. Consequential decisions include those related to employment, financial services, housing, insurance, government benefits, healthcare, legal services, and education.

    The law imposes two separate sets of obligations:

    For Developers

    • Exercise reasonable care to avoid algorithmic discrimination
    • Provide deployers with a statement of dependencies documenting the system’s intended uses, known limitations, and data governance practices
    • Conduct and document impact assessments of the system
    • Make disclosures available to the Colorado Attorney General upon request

    For Deployers

    • Exercise reasonable care to avoid algorithmic discrimination
    • Conduct an impact assessment before deploying the system
    • Implement a risk management plan — which may be based on the NIST AI Risk Management Framework (NIST AI RMF 1.0, published January 2023)
    • Notify consumers when they are interacting with an AI system (where applicable)
    • Provide consumers with an adverse action notice if an AI system was a substantial factor in a denial decision

    Enforcement

    The Colorado Attorney General has exclusive enforcement authority. There is no private right of action under the law. Violations are treated as deceptive trade practices under the Colorado Consumer Protection Act. The AG may also require developers and deployers to provide documentation of their impact assessments and risk management programs.

    How It Compares

    Colorado’s law is modeled on the EU AI Act (Regulation 2024/1689), which similarly categorizes AI systems by risk level and imposes obligations on high-risk systems. However, unlike the EU AI Act, Colorado’s law does not ban any AI uses outright and provides no private right of action.

    It also differs from New York City Local Law 144, which regulates automated employment decision tools (AEDTs) specifically and requires annual independent bias audits — a requirement Colorado’s law does not include.

    Timeline

    • May 17, 2024: Signed by Governor Polis (who expressed concerns about competitiveness and signed with a statement urging future amendments)
    • February 1, 2026: Law takes effect

    Primary Sources


    AI automated and generated content. Human reviewed. All policy text is summarized from the primary sources linked above.