New York will require AI developer registration and restrict public-sector automation; California orders safety recommendations
AI oversight is moving from broad policy debate toward concrete state requirements, agency planning and congressional proposals, but many measures remain under development rather than in force. New York is pairing developer registration with stronger public-sector employment protections, while California is studying monitors and emergency shutdowns; in Washington and Silicon Valley, leaders remain divided over whether voluntary safety work or binding controls should come first.
Monday, Sep 21, 2026
Tracking: AI Legislation · AI law · AI regulation
New York advances AI oversight while limiting public-sector automation
New York is moving from AI policymaking to implementation. Gov. Kathy Hochul is requiring major AI developers to register with the state beginning in November and has announced the first hire for the office overseeing the RAISE Act.
At the same time, a separate state law, the LOADinG Act, is creating a potential constraint on government automation.
The New York Post says amendments shifted the law away from its original transparency focus and now bar public employers from using AI to displace workers, reduce hours, pay or benefits, or transfer duties to automated systems.
The protections reached New York City employees in February 2026, leaving agencies uncertain how broadly the rules apply.
Together, the developments put developers under a new registration regime while public agencies face a more protective employment framework.
The Post links that framework to possible disputes over routine automation, though the law’s key terms remain undefined.
Key facts:
Hochul will require major AI developers to register with New York beginning in November.
The state announced the first hire for the office overseeing the RAISE Act.
LOADinG protections reached New York City employees in February 2026.
The law covers public employers, including school districts, SUNY, CUNY and local governments.
Its key terms, including “displacement,” remain undefined.
Why it matters: New York is pursuing two different forms of AI governance at once: developer registration under the RAISE Act and employment protections restricting automation in public agencies.
The first creates a new compliance step for major AI companies; the second gives public-sector employees and unions a stronger basis to challenge systems that affect existing work.
The main uncertainty is how agencies will interpret the LOADinG Act’s undefined terms.
The New York Post argues that managers may avoid automating routine tasks such as intake, triage, drafting and document review, limiting potential savings as New York City confronts a projected $7.25 billion budget gap.
The next developments to watch are the November registration process and legal or labor disputes over the law’s scope.
Newsom Orders California AI Safety Plan With Kill Switch Proposal
California Gov. Gavin Newsom signed an executive order on September 18 directing experts and state agencies to develop stronger AI safety and security recommendations.
The proposals may include independent monitors for frontier AI companies, safety plans, and emergency “kill switches” for shutting down models; recommendations are due by November 16.
The move begins a policy process rather than imposing those requirements immediately.
It also exposes divisions in California’s governor’s race: Republican Steve Hilton supports studying external monitoring but calls kill switches a gimmick, while Democrat Xavier Becerra promises firm guardrails without endorsing specific measures.
Newsom’s order follows his 2025 law requiring large AI developers to disclose catastrophic-risk management and report qualifying safety incidents, but he previously vetoed legislation resembling the new proposals.
Key facts:
Newsom’s September 18 executive order convenes experts to develop California AI safety and security guidelines.
Recommendations are due by November 16, potentially enabling a special legislative session before Newsom leaves office.
Proposals include independent monitors, developer safety plans, and emergency shutdown mechanisms for frontier AI models.
Steve Hilton called kill switches a gimmick; Xavier Becerra promised firm guardrails without endorsing specifics.
California’s SB 53 requires large model developers to disclose catastrophic-risk management and report qualifying incidents.
Why it matters: The order could give California a framework for regulating the most advanced AI systems, but it does not itself create a kill-switch requirement or other new binding rule.
The immediate consequence for developers is uncertainty: external evaluations, safety plans, and emergency shutdown capabilities could become compliance obligations if lawmakers adopt the recommendations.
The November deadline creates a near-term test for Newsom and the next governor.
A special legislative session could accelerate action, while disagreement among candidates over specific safeguards signals that implementation—and the balance between safety oversight and keeping AI companies in California—remains unsettled.
House lawmakers pair AI-agent controls with national-security reviews
Representatives Josh Gottheimer and Mike Lawler used a September 20 press conference to promote the bipartisan Stop Rogue AI Act, aimed at AI agents operating inside business and government networks.
The bill would direct NIST to develop national standards for finding, verifying, monitoring, and controlling those agents, including the ability to cut off access when necessary.
On September 18, Gottheimer announced a broader AI-safety agenda that included the American AI Security Act and the China FIREWALL Act.
The first would require developers to give the NSA access to covered powerful models for pre-deployment national-security reviews lasting 30 days, with one possible extension; the second would bar Chinese-developed open-weight models from federal devices and related procurement.
Key facts:
Gottheimer and Lawler promoted the Stop Rogue AI Act on September 20, 2026.
The bill would direct NIST to set standards for discovering, verifying, monitoring, and controlling AI agents.
The American AI Security Act would require NSA review before covered models are released.
Reviews would last 30 days, with one possible 30-day extension.
China FIREWALL would bar Chinese-developed open-weight models from federal devices and related procurement.
Why it matters: Together, the proposals target two control points: visibility after AI agents enter networks and screening before powerful models are deployed.
NIST standards could give organizations a common way to identify and disable unauthorized agents, while NSA reviews would add a national-security checkpoint for covered models.
The measures are proposals, not enacted requirements, so their immediate effect is to define a legislative agenda.
Developers could face new testing, access, and review obligations if Congress advances them; federal agencies would face tighter limits on Chinese open-weight software.
Key next steps are congressional action, the scope of models covered, and whether NIST or the NSA receives implementation authority.
Trump Resists Tech Leaders’ Calls for Stronger AI Oversight
Calls from prominent technology executives for stronger AI oversight are colliding with resistance from President Donald Trump and congressional inaction.
On September 15, PBS reported that Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman and xAI’s Elon Musk are among leaders seeking greater scrutiny, while Trump dismissed efforts to limit AI as a “conspiracy.
” Democrats are trying to raise pressure: House members met privately Tuesday, Sen. Bernie Sanders scheduled an expert briefing Wednesday, and candidates including Sen. Jon Ossoff are advocating sweeping action.
But Speaker Mike Johnson has deferred to the White House, and Republicans including Vice President JD Vance have expressed caution; Trump also rescinded the Biden administration’s 2023 executive order on AI oversight, while an earlier bipartisan safeguards recommendation received little follow-up.
Key facts:
Trump called efforts to limit AI a “conspiracy” on Monday.
Amodei, Altman, and Musk are associated with calls for greater government oversight.
House Democrats held a private AI discussion Tuesday.
Sanders scheduled an expert AI briefing for Wednesday.
Hawley opened an investigation into OpenAI’s system hacking another AI company.
Why it matters: For AI developers, the immediate federal landscape remains largely unchanged: Congress has not imposed new rules, and Trump rescinded the prior administration’s executive order.
The result is a widening gap between warnings from technology leaders and the government’s ability to establish safeguards, a gap the article links to partisanship and election-year pressure.
The next test is whether Democratic briefings, Hawley’s investigation, or unusual overlap between Sanders and Trump ally Steve Bannon produces legislation rather than more warnings.
Previous efforts—including a bipartisan recommendation for $32 billion in AI investment and safeguards—received little follow-up, suggesting that attention alone has not translated into federal policy.
Switzerland Plans AI Rules While Existing Law Governs Legal Work
Switzerland is preparing a consultation draft on artificial-intelligence regulation, with the Federal Office of Justice expected to complete it by the end of 2026.
The prospective measures would address transparency, data protection, non-discrimination and supervision, but the government has not enacted an overarching AI statute. For law firms, the immediate rules come from existing law.
The Federal Act on Data Protection applies when AI processes personal data; lawyers’ duties of care and professional secrecy continue when outside providers are involved, alongside possible contract, employment, copyright and procedural obligations.
The practical test is the workflow, not the “AI” label: teams should examine purpose, data flows, access, retention, subcontractors, foreign disclosures and whether processing creates a high risk requiring a data-protection impact assessment.
They should also scrutinize automated decisions affecting individuals and avoid treating the future consultation or international instruments as current duties.
Key facts:
Switzerland has no overarching AI statute based on the EU model.
Federal Office of Justice is expected to complete an AI consultation draft by end-2026.
Proposed measures target transparency, data protection, non-discrimination and supervision.
The Federal Act on Data Protection applies when AI systems process personal data.
Professional secrecy remains applicable when external AI providers handle legal work.
Why it matters: Swiss legal teams can assess AI under established rules rather than pause for a new act, but that flexibility shifts responsibility onto workflow-level controls.
Privacy, confidentiality and professional-care reviews become practical gatekeepers, especially where sensitive files, third-party processing or cross-border disclosure are involved.
The main risk is regulatory overreach in both directions: assuming every AI use requires a data-protection impact assessment, or assuming no AI Act means no obligations.
The next concrete milestone is the consultation draft; until it appears, firms should monitor the proposal and separately assess whether their role, system or connection with the EU creates additional requirements.
Frontier AI Leaders Push Voluntary Pacing and Earlier Safety Checks
Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman are urging frontier AI companies to slow capability gains enough for safety work to keep pace, rather than waiting for federal or international mandates.
Their proposals stop short of halting development: Amodei backs embedded third-party evaluators, coordination among companies in democratic states, and eventual international cooperation, while Altman supports shared standards and independent audits.
OpenAI has also moved safety assessments earlier, requiring explicit safety cases before reinforcement-learning runs expected to produce significant capability gains.
The proposals arrive as no binding international regime limits AI development speed, and the Council of Europe’s AI treaty had not entered force by September 2026.
Any binding pacing system would need precise thresholds, verification, and rules balancing commercial secrecy with national security.
Key facts:
Dario Amodei’s September 2026 proposal calls for slowing frontier-model capability advances.
Sam Altman says OpenAI now prepares safety cases before significant reinforcement-learning runs.
Amodei proposes embedded third-party evaluators and coordination among frontier companies in democratic states.
No binding international regime currently limits AI capability growth, training compute, or development speed.
The Council of Europe AI treaty had not entered force by September 2026.
Why it matters: The convergence between Anthropic and OpenAI leadership could help establish common safety practices before governments impose formal requirements.
Earlier safety cases also give safety teams influence before potentially important capability gains, rather than only before deployment. Voluntary restraint remains difficult when companies and states fear losing commercial or strategic advantages.
The next test is whether frontier labs accept shared standards and external audits, and whether governments can create reciprocal rules with clear thresholds and credible verification.
AI Safety Debate Splits Washington and Silicon Valley
An open dispute over AI regulation has crystallized around competing views of safety, speed and government’s role.
On September 16, CNBC reported President Donald Trump, Nvidia CEO Jensen Huang and David Sacks opposing tighter intervention, while Anthropic’s Dario Amodei, OpenAI’s Sam Altman, Elon Musk and researchers backed slowing frontier-model development.
Huang said at Salesforce’s Dreamforce conference that “safety is an engineering problem, not a legal one,” arguing companies and market pressure can manage risks without new laws.
The opposing camp gained momentum after an Anthropic researcher’s warning and an alignment lead’s claim that AI has a greater than 10% chance of killing all humans; Speaker Mike Johnson said Trump plans to bring AI leaders to the White House within the next week.
Key facts:
Jensen Huang said companies need no new AI laws or regulations at Salesforce’s Dreamforce conference.
Dario Amodei called for slowing frontier-model development over concerns about AI risks.
Trump called fears of runaway AI a “hoax” and “scam” in posts this week.
Anthropic’s alignment lead said AI has a greater than 10% chance of killing all humans.
Mike Johnson said Trump plans to convene AI leaders at the White House within a week.
Why it matters: The dispute is becoming a visible political and commercial divide.
Trump administration figures and Nvidia are presenting rapid development and data-center construction as essential to competing with China, while frontier-lab leaders and researchers argue that development should slow until risks are better controlled.
The immediate policy question is whether the administration’s planned White House meeting produces a common position or exposes the divide further.
Huang’s approach would leave safety largely to companies, markets and existing liability law; the opposing camp is pressing for stronger intervention, while industry self-regulation remains an unresolved alternative.