Public Policy & Governance
Top Line
California Governor Newsom signed an executive order directing state agencies to explore new AI rules including a potential 'kill switch' mechanism, the most concrete regulatory action from a U.S. state executive this week, though it initiates study rather than imposing binding obligations.
Rep. Josh Gottheimer unveiled two bipartisan AI safety bills in the House, a rare cross-party legislative move that signals growing Congressional appetite for federal action even as the Trump administration resists federal AI regulation.
A coalition of both Republican and Democratic states is actively defying the Trump administration's hands-off AI posture, with calls for binding limits accelerating since July in what is shaping into a structural federalism conflict over AI governance.
The UK's AI governance agenda has stalled under Prime Minister Andy Burnham, with senior officials warning the country is losing ground on AI safety legislation that Starmer's government had begun drafting before leaving office.
Europe's absence from the frontline AI safety debate — framed starkly by ECB President Christine Lagarde as a binary choice between exclusion and dependency — underscores how the EU AI Act's consumer-protection framing has left the continent without credible tools to address systemic AI risk.
Key Developments
Newsom Executive Order: Exploratory Action, Not Binding Rule
Governor Newsom's executive order, signed September 18, directs California state agencies to explore new AI regulatory frameworks and evaluate the feasibility of an AI 'kill switch' — a mechanism to halt or constrain frontier model operation in an emergency. Policy professionals should note precisely what this is: an administrative directive to study and report, not a regulation with legal force. Newsom vetoed SB 1047, California's most ambitious AI safety bill, in 2024, making this EO a politically significant recalibration but not yet a reversal of that posture. The order does not impose compliance obligations on AI developers, set timelines for rulemaking, or define the legal authority under which a kill switch would operate. Politico
The strategic significance of the EO lies in its signaling function within the U.S. federal-state standoff. California controls the largest domestic AI market by company concentration, and its regulatory posture shapes industry expectations nationally. If this EO leads to a formal rulemaking process at the California Department of Technology or a new statutory proposal, it will matter enormously. If it produces only a report, it joins a long queue of AI review exercises that generated little enforceable output.
Federal-State AI Governance Fracture Deepens
The bipartisan state defiance of Trump's federal AI non-intervention stance — now spanning both Republican and Democratic administrations at the state level — represents a structural governance problem that no single legislative fix will resolve. According to Politico, calls for more robust state-level limits have escalated since July, suggesting a triggering event — likely a combination of publicized AI harms and the absence of federal action — drove coordination among state legislators across party lines. This is not routine partisan friction; it is a cross-ideological consensus that federal inaction is itself a policy choice states must compensate for.
Into this environment, Rep. Gottheimer's two bipartisan House bills arrive as a potential federal counterweight. Gottheimer sits on the House Democratic Commission on AI and has a track record of brokering cross-aisle deals on financial regulation, giving these bills marginally more credibility than typical opposition-party proposals. The bills are described as targeting risks posed by advanced AI models, but no bill text has been published as of this briefing date, making substantive assessment impossible. The critical question is whether they attract Republican co-sponsors with committee seniority — without that, they will not advance in the current House. Politico
UK AI Governance Vacuum: Institutional Drift at a Critical Juncture
Reporting from The Guardian reveals that the UK's AI safety legislative agenda — which Starmer's government had actively developed, including an internal review of existing statutory powers and preliminary work on compelling frontier AI developers to comply with safety obligations — has effectively stalled under Burnham. Senior figures briefed on the plans are explicitly concerned the issue has dropped off the government's priority list as domestic fiscal and social pressures dominate the agenda. This is a significant institutional setback: the UK AI Safety Institute under Starmer had positioned itself as the leading intergovernmental body for frontier AI evaluation, and legislative authority to back that role was the natural next step.
The practical consequence is that the UK's international influence on AI governance norms now rests on a DSIT-based institute without statutory backing, operating in an increasingly competitive environment where the EU AI Act is in enforcement phase and U.S. states are legislating independently. Without primary legislation, the UK cannot compel model evaluations, mandate incident reporting, or enforce safety standards — it can only request voluntary cooperation from developers, a position that weakens with each passing month as norms harden elsewhere.
Europe's AI Safety Governance Gap: The Act Was Never Designed for This
Christine Lagarde's framing — cited in The Guardian — captures a structural problem the EU AI Act was not designed to solve. The Act is fundamentally a market regulation instrument, built around risk classification of AI applications and conformity assessment obligations for deployers. It does not provide mechanisms to address systemic or existential risks from frontier models, which are developed outside EU jurisdiction and enter European markets as finished products. The AI Office in Brussels, created to coordinate enforcement, lacks the technical capacity and legal tools to evaluate frontier model capabilities in the way the UK AISI or U.S. evaluators can.
The geopolitical dimension compounds this. Europe is dependent on U.S. and increasingly Chinese AI infrastructure at the model layer, and has no comparable domestic frontier model developer to use as a regulatory counterparty. This means EU enforcement actions — when they come — will be reactive, applied to deployed applications rather than upstream model development. The AI Act's phased implementation timeline means full enforcement of the most demanding requirements for general-purpose AI models is still pending, leaving a window where nominal regulation exists but practical leverage does not.
Signals & Trends
The 'Kill Switch' Concept Is Entering Formal Policy Discourse — And Will Force Precision
Newsom's executive order is the first instance of a major executive in a leading AI jurisdiction directing formal study of an AI kill switch mechanism. The concept has circulated in academic and advocacy literature, but its entry into official rulemaking exploration will force policy actors to specify what it actually means: a physical data center cutoff, a model weight deletion requirement, a remote deactivation API mandate, or an emergency licensing revocation power. Each option has radically different technical feasibility, constitutional implications, and industry compliance costs. The definitional work that follows Newsom's EO — assuming it proceeds — will either produce a credible governance tool or expose the concept as technically unworkable, and that determination will ripple through every other jurisdiction considering similar powers.
AI Safety Framing Is Shifting From Voluntary Cooperation to Legislative Compulsion — Unevenly
Across the UK, U.S. states, and California specifically, the policy debate has moved from whether to legislate AI safety to how. The Starmer government was exploring powers to compel frontier developers to cooperate with safety evaluations; Gottheimer's bills target model risks directly; state legislatures are drafting binding rules. The common thread is that the voluntary cooperation model — which undergirded the Bletchley Declaration, the White House voluntary commitments of 2023, and the AI Safety Institute's evaluation program — is no longer treated by legislators as sufficient. This shift is significant because it changes the negotiating position of AI developers from willing participants shaping light-touch frameworks to regulated entities managing compliance exposure, with downstream effects on where companies locate operations, disclose capabilities, and structure lobbying priorities.
Public Sector AI Deployment Is Generating Governance Evidence the Policy Cycle Has Not Caught Up To
The Australian case of Microsoft Copilot advising an MP to congratulate a constituent on a terminal illness decision is emblematic of a broader pattern: governments are deploying AI tools in constituent-facing and decision-support roles before governance frameworks for those deployments exist. The incident, aired at a parliamentary inquiry alongside testimony from the cyber security chief about AI's necessity for threat defense, illustrates the tension that procurement decisions are running ahead of accountability frameworks. Most jurisdictions have general-purpose AI use policies for civil servants, but these rarely address the specific risk profile of AI tools in politically sensitive constituent interactions, emergency response contexts, or national security adjacent functions. The gap between deployment velocity and governance specificity is widening.
Explore Other Categories
Read detailed analysis in other strategic domains