The Race to Set the Rules

The law has a speed problem. Artificial intelligence makes that problem unusually visible because capability can change on a software timetable while public authority still moves through hearings, votes, notice, comment, litigation, and appeal.

A legislature can spend a year defining an artificial intelligence system while the systems themselves change several times. A regulator can finish guidance for one generation of models just as agents begin doing things the guidance never contemplated. A court can spend years deciding who was responsible for a failure that happened in seconds.

The mismatch does not make government irrelevant. It makes the design of governance more important, because the rule has to survive long enough to matter without pretending that a legislature can predict every technical turn.

The central question is no longer whether AI should be regulated. Governments are already regulating it. The harder question is how to write rules that remain useful when the technology, the market, and the evidence keep moving.

Three Clocks Are Running

AI governance operates on at least three clocks. The capability clock measures what systems can do. The legislative clock measures how long institutions take to debate, enact, and revise rules. The enforcement clock measures when those rules actually change behavior.

These clocks can move at different speeds. Passing a law does not mean the people enforcing it already have the staff, tests, or procedures they need. By the time a rule takes effect, the product may have changed. To judge progress, ask separately what the technology can do, what the law requires, and what authorities can actually enforce.

The European Union’s AI Act makes the problem visible. The Act entered into force in 2024. Prohibited practices and AI-literacy duties began applying in 2025. General-purpose AI obligations also began in 2025, with Commission enforcement powers becoming active in August 2026. Transparency duties for chatbots and synthetic content apply as of August 2, 2026; Chapter Nineteen examines the synthetic-content side of these duties. Yet high-risk rules for areas such as employment, education, biometrics, migration, and critical infrastructure have been pushed to December 2027, while rules for AI embedded in regulated products extend to August 2028.

That staggered architecture is not necessarily a failure. Some obligations need standards, guidance, testing methods, and competent regulators before enforcement can be credible. But it demonstrates the governing problem: a statute is a starting point, not a completed control system.

A rule that cannot be updated is a snapshot trying to govern a movie.

The United States Is Running Several Races at Once

The United States has chosen a different path. Instead of one comprehensive federal AI statute, rules are taking shape through presidential policy, federal agencies, laws for particular industries, state laws, and court cases. The continuing argument is partly about federal preemption: when federal law overrides state law, and which AI rules should be left to the states.

The White House’s 2025 AI Action Plan emphasized innovation, infrastructure, and international leadership while directing agencies to reduce rules considered unnecessarily burdensome. A December 2025 executive order then called for a minimally burdensome national AI framework, created an AI Litigation Task Force to challenge certain state laws, and directed the administration to develop federal legislative recommendations that would preempt conflicting state AI regulation.

The same order preserved important categories for state action, including child safety, state procurement and use of AI, and data-center siting. The administration’s March 2026 legislative recommendations kept that structure: a national framework with broad preemption, but room for traditional state powers involving children, fraud, consumer protection, infrastructure, and a state’s own use of AI.

That is more than a jurisdictional fight. It is a disagreement about where accountability should sit. A single national rule can reduce conflicting compliance burdens and prevent one state from effectively writing policy for the country. State experimentation can also expose harms, test remedies, and move faster when Congress does not.

Uniformity can reduce friction. It can also freeze the wrong rule at national scale.

The States Have Become the Laboratory

The state systems already differ enough to show several competing theories of AI governance.

California’s SB 53 focuses on large frontier developers. It requires covered developers to publish frontier AI frameworks describing how they incorporate standards and best practices, creates reporting for critical safety incidents, protects certain whistleblowers, and authorizes civil penalties. It also requires the California Department of Technology to recommend updates annually. That last feature may be as important as the original requirements because it treats technological change as part of the statute’s operating environment.

New York’s RAISE Act, effective January 1, 2027, also targets large frontier developers. It requires published safety frameworks and incident reporting and creates an oversight role within the Department of Financial Services. California and New York therefore converge on a basic idea: the developer of a powerful model should have written safety practices before a catastrophic event forces everyone to ask what the practices were.

Texas has taken a broader but generally lighter-touch route. Its Responsible Artificial Intelligence Governance Act took effect in January 2026 and regulates specified uses of AI, including disclosure in certain settings and prohibitions involving manipulation, unlawful discrimination, constitutional rights, and some biometric uses. Colorado rewrote its consequential-decision law in 2026, with the new automated-decision framework and chatbot rules scheduled to take effect in January 2027.

These laws do not add up to a national system. They add up to experimentation, with different states testing different answers to who must disclose, who must assess risk, which uses are prohibited, and which institutions may enforce the boundary.

The patchwork imposes real compliance costs, especially on smaller firms that cannot maintain fifty legal teams. It also produces information about which rules can be implemented, which create unintended incentives, and which harms are important enough to justify national treatment.

Child Safety Shows Why the Patchwork Persists

Chapter Seventeen described California’s September 2026 child-safety package. The laws strengthened companion-chatbot safeguards, required child-safety risk assessment and auditing, and restricted addictive features for younger users. Together with California’s 2025 age-bracket legislation, the state’s broader child-safety architecture also seeks to support age assurance while reducing unnecessary identity collection.

That package also demonstrates why federal preemption becomes difficult in practice. Child safety is one of the areas the White House itself excluded from its proposed preemption framework. The political system is therefore already acknowledging that some AI harms may need local or state action even while national policymakers argue that model development should not be governed by fifty separate regimes.

The useful distinction is between regulating the model and regulating the consequence. A frontier model may cross every state border. A school, medical board, insurer, employer, court, or child-facing service still acts inside an institutional and legal setting that already has accountable authorities.

Governance becomes more durable when it attaches rules to the decision being made, the risk being created, and the institution that can actually intervene.

Sector Rules Can Be Smarter Than AI Rules

Some of the strongest governance may come from institutions that already understand the consequence.

Medicine provides a good example. The FDA’s 2026 clinical-decision-support guidance does not attempt to regulate every use of AI in health care through one definition. It distinguishes software functions that fall within medical-device oversight from functions excluded by statute or subject to enforcement discretion. The boundary depends on what the software does and whether a health professional can independently review the basis of a recommendation.

In mid-2026, the Washington Medical Commission was considering a different problem: whether AI can be licensed to practice medicine and how accountability should be assigned when physicians use AI tools. The American Medical Association supported the proposed conclusion that AI cannot hold a medical license and that a physician must retain final authority over care. It objected, however, to language that would place full accountability on the physician even when an AI tool failed or was mandated by an employer, health system, or payer.

That dispute could have been written from the Delegation Doctrine. Principle 1 says accountability does not transfer to the machine. It does not say every consequence belongs to the human standing closest to the output. Principle 6 says consequences should reach the owner. If a developer controls a hidden failure mode, a hospital mandates the tool, and a physician has limited ability to detect the defect, assigning everything to the physician creates ownership without control.

Good regulation follows the decision chain rather than the marketing category.

Too Specific, Too Vague, or Built to Adapt

AI rules can fail in opposite directions. A rule written around a narrow technical feature may become obsolete when the feature changes. A rule written only around broad values can survive forever while giving regulated parties no usable instruction about what compliance requires.

A more durable rule separates the stable boundary from the changing measurement. For example, the duty to provide meaningful human review can remain while the tests for adequate review improve. A law can require incident reports while regulators revise what information those reports must contain. The obligation stays recognizable; the way compliance is checked can respond to better evidence.

This also limits a familiar form of regulatory theater. A company can comply perfectly with a stale checklist while violating the purpose the checklist once served. The opposite failure occurs when an agency announces a broad principle but leaves every important term undefined until enforcement. Both weaken legitimacy because one rewards formal compliance without safety and the other makes the boundary visible only after someone crosses it.

A good AI rule should therefore be specific about accountability, prohibited outcomes, required evidence, appeal, and enforcement, while remaining adaptable about the technical methods used to satisfy those duties.

Rules Need an Update Mechanism

The most important feature of an AI law may eventually be the mechanism for changing it.

California’s frontier law requires annual recommendations for statutory updates. The European Commission can revise implementation guidance and periodically update its general-purpose AI Code of Practice. The United Nations has created an Independent International Scientific Panel on AI that is expected to produce recurring assessments for the Global Dialogue on AI Governance.

These mechanisms differ in legal force, but they share an important design principle. The evidence layer should be allowed to move faster than the constitutional layer of the rule.

A durable law should define stable objectives, accountable institutions, prohibited outcomes, rights of appeal, reporting duties, and enforcement authority. Technical thresholds, evaluation methods, model categories, and implementation guidance may need a faster revision cycle.

This is the same architecture used by this Living Edition. The argument should not change every time the news changes. The evidence must be able to.

The update mechanism is therefore part of the rule. A governance system that can learn from incidents, audits, court decisions, technical evaluations, and changed capabilities has a better chance of remaining legitimate than one that treats enactment as the end of governing.

Enforcement Is Where Governance Becomes Real

Passing a law is politically visible. Building an enforcement system is slower and less glamorous, yet enforcement determines whether the statutory boundary becomes an operating fact or merely another PDF announcing good intentions.

The European Commission began enforcing AI Act rules in August 2026 with complaint, whistleblower, and downstream-provider channels alongside the new transparency requirements. Colorado’s attorney general is writing rules before its new automated-decision law takes effect in 2027. New York has created a state oversight structure for frontier developers. California’s SB 53 gives its attorney general civil enforcement authority.

Each example points to the same problem. A rule needs someone who can receive evidence, investigate, demand records, interpret the boundary, impose consequences, and revise practice. Otherwise the law may describe acceptable behavior without creating a credible path from violation to consequence.

Principle 6 applies to governments too. Consequences must be able to reach the owner.

A prohibition without an enforcement path is a preference written in legal type.

Spain’s September 2026 position makes the institutional point directly: artificial intelligence cannot be governed only by those who control and profit from the technology. Prime Minister Pedro Sánchez framed AI oversight as a public responsibility, tied not only to model behavior but also to cybersecurity, data autonomy, energy use, environmental standards, and the communities that host the infrastructure. That combination matters. A frontier model is not only a product. It is part of a social and physical system. When deployment depends on data centers, public power, national security, and public trust, self-regulation cannot carry the full accountability burden. The accountable chain must include institutions with authority outside the company itself.

The Regulator Needs Expertise Too

Chapter Eighteen argued that verification becomes more valuable as generation becomes cheaper. Regulators face the same verification premium. An agency cannot supervise frontier systems effectively if it lacks the technical competence, access, testing capacity, and institutional authority to challenge the companies it regulates.

Government does not need to employ more AI researchers than industry. It does need people capable of questioning a company’s evidence, spotting missing information, and recognizing a test designed to make the product look good. Outside auditors can help. But if an agency cannot judge the auditors’ work, it has handed away another layer of the oversight it was meant to provide.

Regulatory competence also has to be maintained. Agencies need access to current expertise, incident data, model evaluations, and cross-border evidence. A regulator whose technical knowledge freezes at the date a statute passed will eventually supervise a system that no longer exists.

Government therefore has its own version of Principle 8: accountability decays when competence decays.

International Governance Has a Different Job

Global institutions face an even harder version of the problem because the technology crosses borders while legal authority does not. A model can be trained in one country, hosted in another, incorporated into a service in a third, and affect a person in a fourth.

The United Nations held its first Global Dialogue on AI Governance in Geneva in July 2026, with more than 4,200 registered participants from nearly 170 Member States. The Dialogue was paired with the new 40-member Independent International Scientific Panel on AI, which released its preliminary report days before the meeting.

The arrangement does not create a world AI regulator. It creates something more modest and still useful: a shared evidence process and a forum where governments can compare approaches. The Panel itself describes the timing problem plainly. Policymakers need evidence before acting, but waiting until evidence is conclusive can make intervention too late.

International coordination matters most where national rules can be evaded by moving development, deployment, compute, or distribution across borders. It also matters where incompatible national rules make compliance impossible or where one country’s failure creates risks for others.

Chapter Twenty-One takes the harder question of red lines, including autonomous weapons and other uses that governments may decide should be prohibited rather than merely managed. This chapter’s narrower point comes first: countries need enough common language and evidence to know what they are trying to govern.

Current Evidence: The Rules Are Still Moving

Reviewed September 19, 2026.

The European Union crossed an important threshold on August 2. The AI Act is now broadly applicable, Commission enforcement powers are active, and transparency requirements for human-AI interaction and synthetic content are in force. At the same time, high-risk obligations have been delayed into 2027 and 2028 under the AI Omnibus political agreement. Europe therefore demonstrates both the strength and the difficulty of comprehensive regulation: the framework exists, but implementation has to keep adapting around it.

The United States remains structurally unsettled. The White House is pursuing a national framework that would preempt some state AI laws, while preserving state authority in areas such as child safety. Congress has not yet enacted that comprehensive framework.

The states continue to move. California’s frontier-model law is in force, New York’s RAISE Act begins in January 2027, Texas’s AI governance law has been effective since January 2026, and Colorado is writing rules for its January 2027 automated-decision and chatbot laws. California also signed a major child-safety package on September 10, reinforcing the state-level role the federal proposal itself would preserve.

On September 18, 2026, California directed accelerated implementation of independent AI verification and auditor oversight. Executive Order N-9-26 also required recommendations by November 16 on possible requirements for embedded evaluators, independently verified safety reports, and emergency shutoffs whose effectiveness would be checked repeatedly. These proposed developer obligations were not yet enacted requirements. The distinction matters: announcing a safeguard does not establish that it works. For institutions delegating consequential work, the practical obligation is to identify who can intervene, test that capacity under realistic conditions, and establish who must act when the test fails. A claimed ability to stop the system is only the beginning of an accountability structure.

On October 28, 2026, Reuters reported that Anthropic had asked California Governor Gavin Newsom to support allowing officials to seek court orders blocking the release of AI models posing a serious public risk. The company’s October 15 letter also argued against required switches intended to shut down entire models, favoring interventions against a specific agent or a customer’s access. These are reported policy recommendations, not enacted powers or evidence that a stopping mechanism works.

An instruction to stop needs a defined target and someone with authority to enforce it. Blocking a release, suspending one agent and withdrawing an account’s access are different interventions. A regulator should require evidence about what each measure can halt, what can continue elsewhere, and how compliance will be checked. Chapter Twenty-Five’s operational stop test therefore needs a public counterpart: who can order intervention, on what evidence, and with what opportunity to challenge the decision.

The federal debate shifted again on September 11. Reuters reported that bipartisan Senate negotiators were discussing a bill that could impose a duty of care on developers of the most advanced models, require testing aimed at catastrophic risks, permit the federal government to block release of some models subject to court challenge, and preempt certain state rules in the same risk area. The proposal remains under negotiation and may never become law. Its importance for this chapter is structural: even the emerging federal approach is wrestling simultaneously with safety duties, judicial review, national uniformity, and state authority.

Industry positions are moving too. OpenAI said on September 9 that it now supports mandatory, capability-based national safety regulation for advanced systems, including independent assessments, cybersecurity protections, and incident reporting, while continuing to support several California bills. The company tied that position in part to the prospect of AI-accelerated AI development, reconnecting the research-acceleration thread from Chapters One, Three, and Eighteen without reopening the full case. OpenAI systems are used in this Living Edition’s production workflow, so that position is recorded as the company’s stated view rather than independent evidence that its preferred framework is the right one.

The United Nations has meanwhile created an annual scientific assessment and dialogue process rather than a binding global code. That choice recognizes a basic constraint: international consensus may move more slowly than the technology, but a common evidence base can still reduce the chance that every government is regulating a different imagined system.

The race to set the rules is therefore not producing one winner. It is producing competing regulatory architectures in real time.

On October 12, 2026, the United Kingdom said its AI Security Institute was conducting rigorous evaluations of the latest generation of advanced models after work with leading developers and the U.S. government. The government also said it would consider stronger protections for serious AI incidents and near misses through the Cyber Security and Resilience Bill regime, and was preparing for incidents involving autonomous AI agents. These are government announcements and proposed measures, not proof that the evaluations or reporting system will be sufficient.

The governance lesson is that oversight requires access before the failure becomes public. A regulator or safety institute cannot evaluate a frontier system it cannot inspect, and incident learning depends on reports that arrive soon enough to matter. For consequential AI, public accountability increasingly depends on pre-release testing access, near-miss records, and authority to respond when an agent behaves unexpectedly.

Accountability Analysis

Principles 1, 3, 5, and 6 carry most of this chapter. Principle 1 prevents governments from treating the existence of a regulator as proof that accountability has been solved. The developer, deployer, institution, professional, and government can each own different parts of the chain.

When an AI agent acts, the legal and moral question should not begin with whether the agent ‘broke loose.’ It should begin with who authorized the tool, who gave it access, who set its objective, who monitored its behavior, and who had the power to stop it.

Principle 3 asks for a boundary. Good law says what a system may do, what requires escalation or review, and what is prohibited. It also identifies the scope of the rule clearly enough that people can know whether it applies before enforcement begins.

Principle 5 is the governing principle for fast systems. Speed and architecture move accountability upstream. When case-by-case review cannot keep pace, law has to govern earlier choices: evaluation before release, reporting after incidents, documentation of safety frameworks, auditability, procurement rules, and the conditions under which a system may enter a consequential domain.

Principle 6 asks whether consequences can reach the actor that actually controlled the risk. A law that assigns liability to the nearest visible human while insulating the developer, employer, agency, or institution that controlled the architecture may create the appearance of accountability while breaking the chain.

Regulation is itself a delegation system. Legislatures delegate detail to agencies. Agencies delegate compliance work to firms and auditors. Firms delegate execution to models and employees. The Accountability Test applies all the way back up.

Follow a rule into practice

Choose one AI rule discussed in this chapter. Identify who must comply, what evidence must be available, and which authority can investigate a failure. Then check its status: proposed, enacted but not yet effective, or already in force. Finally, ask how it can be updated. This shows how far the rule has traveled from announcement to practical control.

Accountability Verdict

The goal should not be to make AI law move as fast as AI. Law earns legitimacy partly by moving more slowly, hearing competing interests, and creating predictable rights. The goal is to make the system of law adaptive enough that deliberation does not become paralysis.

Stable principles should move slowly. Technical rules should move faster. Evidence should move faster still.

The best governance will combine national consistency where scale demands it, sector expertise where consequences differ, state experimentation where local authority matters, and international coordination where risks cross borders. None of those layers eliminates the others.

After the technology changes, a rule must still answer the Accountability Chain’s five questions: Who is accountable? What is the boundary? What evidence must be visible? Who can intervene? Where do the consequences land?

The technology will keep changing. The accountability chain cannot be allowed to disappear between versions.

Chapter Twenty-One asks where adaptation should stop. Some decisions may require a boundary that does not move at all.

Sources and Notes

1. European Commission, “AI Act,” current implementation timeline; and “Commission starts enforcing AI Act rules and new transparency requirements on 2 August,” July 31, 2026. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai and https://digital-strategy.ec.europa.eu/en/news/commission-starts-enforcing-ai-act-rules-and-new-transparency-requirements-2-august. Accessed September 12, 2026. The Commission states that the AI Act became broadly applicable on August 2, 2026, with earlier application for prohibited practices, AI literacy, governance, and general-purpose AI obligations. Following the AI Omnibus political agreement, Annex III high-risk rules are scheduled for December 2, 2027 and product-embedded high-risk rules for August 2, 2028. Commission enforcement powers and new transparency duties began applying August 2, 2026. Source record: DA-CH20-S1. Archive reference: pending.

2. The White House, Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence,” December 11, 2025; and “National Policy Framework for Artificial Intelligence: Legislative Recommendations,” March 20, 2026. https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/ and https://www.whitehouse.gov/wp-content/uploads/2026/03/03.20.26-National-Policy-Framework-for-Artificial-Intelligence-Legislative-Recommendations.pdf. Accessed September 12, 2026. The executive order directs creation of an AI Litigation Task Force, evaluation of state AI laws, and development of a national framework. The legislative recommendations call for federal preemption of some state AI laws while preserving state powers in areas including child protection, fraud, consumer protection, infrastructure siting, and state government use. Source record: DA-CH20-S2. Archive reference: pending.

3. The White House, “America’s AI Action Plan,” July 2025. https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf. Accessed September 12, 2026. The plan identifies more than 90 federal policy actions across innovation, infrastructure, and international diplomacy and security, and calls for reducing federal and state regulatory barriers judged to impede AI development while acknowledging a role for prudent state laws. The chapter uses the plan as the administration’s stated policy framework, not as an independent assessment of regulatory effects. Source record: DA-CH20-S3. Archive reference: pending.

4. Office of Governor Gavin Newsom, “Governor Newsom signs SB 53, advancing California’s world-leading artificial intelligence industry,” September 29, 2025. https://www.gov.ca.gov/2025/09/29/governor-newsom-signs-sb-53-advancing-californias-world-leading-artificial-intelligence-industry/. Accessed September 12, 2026. California’s Transparency in Frontier Artificial Intelligence Act requires covered frontier developers to publish safety frameworks, establishes critical-safety-incident reporting and whistleblower protections, authorizes civil penalties, creates CalCompute, and directs the California Department of Technology to recommend updates annually. Source record: DA-CH20-S4. Archive reference: pending.

5. Office of Governor Kathy Hochul, “Governor Hochul Signs Nation-Leading Legislation to Require AI Frameworks for AI Frontier Models,” December 19, 2025; New York General Business Law Article 44-B, current revision. https://www.governor.ny.gov/news/governor-hochul-signs-nation-leading-legislation-require-ai-frameworks-ai-frontier-models and https://www.nysenate.gov/legislation/laws/GBS/A44-B. Accessed September 12, 2026. New York’s RAISE Act requires large frontier developers to maintain and publish frontier AI frameworks and report specified incidents, creates oversight within the Department of Financial Services, and is effective January 1, 2027. Source record: DA-CH20-S5. Archive reference: pending.

6. Texas Legislature, House Bill 149, Texas Responsible Artificial Intelligence Governance Act, enrolled bill summary and text, effective January 1, 2026. https://capitol.texas.gov/billlookup/BillSummary.aspx?Bill=HB149&LegSess=89R and https://capitol.texas.gov/tlodocs/89R/billtext/html/HB00149F.HTM. Accessed September 12, 2026. The law creates disclosure requirements in specified contexts and prohibits certain uses involving manipulation, biometric data, constitutional rights, and unlawful discrimination. The chapter uses Texas as an example of a state framework materially different from frontier-model safety laws in California and New York. Source record: DA-CH20-S6. Archive reference: pending.

7. Colorado Attorney General, “Colorado Automated Decision-Making Technology & Chatbot Safety Rulemaking,” current edition. https://coag.gov/ai/. Accessed September 12, 2026. The Colorado Attorney General states that SB 26-189 repealed and reenacted the state’s earlier high-risk AI provisions as a new automated-decision framework effective January 1, 2027, and that HB 26-1263 created chatbot-safety requirements effective the same date. The office is conducting rulemaking before implementation. Source record: DA-CH20-S7. Archive reference: pending.

8. Office of Governor Gavin Newsom, “Governor Newsom signs the strongest child safety chatbot and social media laws in the nation,” September 10, 2026. https://www.gov.ca.gov/2026/09/10/governor-newsom-signs-the-strongest-child-safety-chatbot-and-social-media-laws-in-the-nation/. Accessed September 12, 2026. The package includes child-facing chatbot safeguards and restrictions on addictive social-media features. Chapter Seventeen contains the detailed child-safety analysis; Chapter Twenty uses the package only to show the continuing role of state regulation and the child-safety exception in the federal preemption debate. Source record: DA-CH20-S8. Archive reference: pending.

9. American Medical Association, “Aug. 28, 2026: State Advocacy Update,” August 28, 2026; and “AMA Principles for Augmented Intelligence Development, Deployment, and Use,” current policy. https://www.ama-assn.org/health-care-advocacy/advocacy-update/aug-28-2026-state-advocacy-update and https://www.ama-assn.org/system/files/ama-ai-principles.pdf. Accessed September 12, 2026. The AMA supported the Washington Medical Commission’s proposed position that AI cannot hold a medical license and that physicians retain final authority over patient care, while objecting to assigning physicians full accountability for AI failures even when a tool is mandated or its defect is outside reasonable clinical detection. AMA policy argues that accountability and liability should align with the parties best positioned to know and mitigate risks. The AMA is an advocacy organization representing physicians; its position is presented as such. Source record: DA-CH20-S9. Archive reference: pending.

10. U.S. Food and Drug Administration, “Clinical Decision Support Software,” final guidance, January 2026; and “Clinical Decision Support Software Frequently Asked Questions,” updated 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software and https://www.fda.gov/medical-devices/software-medical-device-samd/clinical-decision-support-software-frequently-asked-questions-faqs. Accessed September 12, 2026. FDA guidance distinguishes software functions subject to medical-device oversight from specified non-device clinical decision support and other functions subject to enforcement discretion. The chapter uses the guidance as an example of sector-specific governance tied to the function and consequence rather than to a universal AI category. Source record: DA-CH20-S10. Archive reference: pending.

11. United Nations, Global Dialogue on AI Governance, 2026 materials; and Independent International Scientific Panel on AI, Preliminary Report and FAQ, July 2026. https://www.un.org/global-dialogue-ai-governance/en and https://www.un.org/independent-international-scientific-panel-ai/en/preliminary-report. Accessed September 12, 2026. The first Global Dialogue was held in Geneva on July 6-7, 2026 with more than 4,200 registered participants from nearly 170 Member States. The 40-member Scientific Panel released its preliminary report on July 1 and is designed to provide recurring evidence-based assessments to the Dialogue. The bodies do not constitute a binding global AI regulator. Source record: DA-CH20-S11. Archive reference: pending.

12. Courtney Rozen, Reuters, “US Senate negotiators consider requiring AI firms to mitigate known major risks,” September 11, 2026; and Chandni Shah, Reuters, “OpenAI pushes for mandatory national AI safety rules,” September 9, 2026. https://www.reuters.com/legal/litigation/us-senate-negotiators-consider-requiring-ai-firms-mitigate-known-major-risks-2026-09-11/ and https://www.reuters.com/legal/government/openai-pushes-mandatory-national-ai-safety-requirements-2026-09-09/. Accessed September 12, 2026. Reuters reports that bipartisan Senate negotiators are discussing, but have not enacted, a framework that could impose a duty of care on developers of the most advanced models, involve federal testing and possible release restrictions subject to judicial challenge, and preempt some state rules concerning catastrophic model risks. Reuters also reports OpenAI’s call for mandatory capability-based national safety rules and its support for several California bills. The Senate proposal remains under negotiation. OpenAI’s position is a provider position; OpenAI systems are used in this Living Edition’s production workflow, so the standing disclosure applies. Source record: DA-CH20-S12. Archive reference: pending.

13. California Executive Department, “Executive Order N-9-26,” September 18, 2026, operative paragraphs 1-3. https://www.gov.ca.gov/wp-content/uploads/2026/09/FINAL-N-9-26-AI-EO-9.18.26-SIGNED.pdf. Accessed September 19, 2026. The order accelerates independent-verification and auditor oversight work and requires recommendations by November 16, 2026 on possible developer obligations, including embedded evaluators, independent verification of safety reports, and repeatedly verified emergency shutoffs. It does not enact those proposed obligations or establish that the safeguards are effective. The passage connects this chapter’s regulatory-competence argument with Chapter Twenty-Five’s controlled intervention exercises. Source record: DA-CH20-S13. Editorial MODIFY 018; formal Update 033. Archive reference: pending.

14. Victoria Waldersee, Benjamín Mejías Valencia, and David Latona, Reuters, “Spanish PM Sanchez says AI industry cannot be self-regulated,” September 21, 2026. https://www.reuters.com/world/spanish-pm-sanchez-says-ai-industry-cannot-be-self-regulated-2026-09-21/. Accessed September 21, 2026. Reuters reports Sánchez’s statement that AI cannot be self-regulated by those who control the technology, and his 12-month roadmap covering responsible deployment, cybersecurity, a proposed AI gigafactory, models for climate, health and energy, environmental and energy standards for data centers, data autonomy, and local community benefit. Source record: DA-CH20-S14. Archive reference: pending.

15. Jody Godoy, Reuters, ‘FTC chair suggests AI developers should be liable for conduct of agents,’ September 25, 2026. Original URL: https://www.reuters.com/business/ftc-chair-pushes-back-treating-ai-agents-independent-actors-2026-09-25/. Verified Reuters syndication: https://www.channelnewsasia.com/business/ftc-chair-suggests-ai-developers-should-be-liable-conduct-agents-6411691. Accessed September 26, 2026. Reuters reports that FTC Chair Andrew Ferguson rejected treating AI agents as independent actors and suggested that developers who instruct them could be liable for harm. This records his stated enforcement position, not an enacted liability rule, a court finding, or proof of wrongdoing by any named company. The accompanying accountability questions are the author’s normative analysis. Source record: DA-CH20-S15. Archive reference: pending.

16. GOV.UK, Cabinet Office, AI Security Institute, National Cyber Security Centre, and Department for Digital, Culture, Media and Sport, “Britain steps up national defences against growing AI risks,” October 12, 2026. https://www.gov.uk/government/news/britain-steps-up-national-defences-against-growing-ai-risks. Accessed October 13, 2026. Government announcement of ongoing AI Security Institute evaluations, consideration of stronger serious-incident and near-miss protections through the Cyber Security and Resilience Bill regime, and preparation for incidents involving autonomous AI agents. These are announcements and proposed measures, not proof of sufficiency or effectiveness. The access-and-reporting accountability lesson is the author’s analysis. Source record: DA-CH20-S16. Update 094. Archive reference: pending.

17. Courtney Rozen, Reuters, “Anthropic tells California governor courts should be able to block risky AI models,” October 28, 2026. https://www.reuters.com/legal/government/anthropic-tells-california-governor-courts-should-be-able-block-risky-ai-models-2026-10-28/ Accessible syndication: https://www.investing.com/news/world-news/anthropic-tells-california-governor-courts-should-be-able-to-block-risky-ai-models-4980201. Accessed October 29, 2026. Independent reporting of an October 15 company letter, authenticated by a company spokesperson; original letter not independently read for this brief. Policy recommendations, not enacted law or evidence of effective intervention. Paragraph two is the author’s analysis. Anthropic systems are used in this Living Edition’s external critique workflow. Source record: DA-CH20-S17. Update 122. Archive reference: pending.