The arc just closed.
For months, the federal AI executive order occupied a central position in the compliance planning calculus of every enterprise with AI exposure. According to reporting by the South China Morning Post, Axios, and the Associated Press, Trump canceled the order hours before its scheduled May 21 signing ceremony. The draft had been described as near-final that same morning. The voluntary pre-release security review framework, a 90-day window requiring frontier model developers to submit systems to federal agencies before public release, won’t take effect. No replacement has been announced.
The cancellation is now a known fact. What that fact means for the compliance landscape over the next 12 months is the question this piece answers.
What the 90-Day Framework Would Have Required
Start with what was killed. The draft executive order would have established a voluntary pre-release security review: frontier model developers submit systems to federal agencies, agencies evaluate them during a window of up to 90 days, and public release holds during that period. Voluntary, not mandatory. No statutory penalties. No enforcement mechanism with teeth.
That’s the lightest possible federal intervention in AI deployment. It wasn’t a ban. It wasn’t a licensing regime. It was a structured checkpoint that asked companies to show their work to the government before releasing systems to the public.
Trump stated, according to SCMP’s reporting on an Oval Office press briefing, that he didn’t want any action that would impede America’s technological lead over China. Former White House tech policy advisor Dean Ball characterized the cancellation as reflecting “healthy tension” within the administration, per the Los Angeles Times. That’s attributed interpretation, not a confirmed account of the internal deliberations.
The practical consequence of the cancellation is this: there’s no federal pre-release framework, voluntary or otherwise, now in the pipeline. OpenAI, Anthropic, and Google were named parties to the policy discussions. The CAISI voluntary agreements remain active, but those are softer commitments, not structured review obligations with defined timelines.
The Stakeholder Map
Understanding why the EO died requires knowing where each actor stood.
The frontier AI labs had a consistent posture: voluntary pre-release review was acceptable in principle, but the specific terms, the 90-day window, the agency review scope, the submission requirements, created competitive exposure. A 90-day delay in public release, even voluntary, creates timing disadvantages in a market where model releases are closely watched. Anthropic’s Head of State and Local Government Relations appeared at the CalMatters Festival in Sacramento on May 21, the day of the cancellation, publicly defending Claude deployment in California state agencies. That appearance, on the same day the federal framework died, illustrates where the frontier lab community was investing its relational capital: state governments, not federal review frameworks.
Dean Ball’s “healthy tension” framing, while an attributed opinion, names something real: the Trump administration contains two competing instincts on AI governance. The competitiveness faction, aligned with the labs, argues that any federal friction costs the U.S. its lead against China. A regulatory-skeptic caucus within the administration was always the structural obstacle to a federal AI review process, voluntary or not. The cancellation is that faction winning.
State AI Regulatory Model Comparison
Who This Affects
The state-level actors weren’t waiting for the federal framework to resolve. California was already mid-process on its workforce EO. Colorado’s S.B. 26-189 was already law. Illinois had introduced an eight-bill AI package earlier in May. These state actions weren’t contingent on the federal outcome, but the federal cancellation removes the one instrument that would have given states reason to pause.
The State Acceleration Response
Three state-level models are now active simultaneously, and they reflect fundamentally different theories of what AI regulation should do.
California’s executive order, signed on May 21, directs state agencies to analyze union CBAs for AI provisions, develop retraining programs, study job subsidies and stock compensation policies, and build a public dashboard tracking AI-related job losses. According to CalMatters reporting, the EO initiates studies and agency reviews, no private employer compliance obligations exist today. But the infrastructure it builds, a displacement tracking dashboard, retraining program capacity, a legal map of CBA AI provisions, is the foundation for enforcement-capable instruments that could follow.
Colorado’s S.B. 26-189 is a different model entirely. It’s a disclosure regime: a private-sector obligation requiring employers using algorithmic decision-making tools in high-stakes contexts to disclose that use to affected individuals. The compliance deadline is January 1, 2027. It creates direct private employer obligations today, with HIPAA exemptions and a federal cap clause that are still being interpreted.
Illinois introduced an eight-bill AI package in May 2026. The package covers a range of AI policy areas. Unlike California’s government-response model and Colorado’s disclosure regime, Illinois’s package represents a legislative ambition to address multiple AI policy problems in a single session, a scope that creates compliance complexity for employers with multistate workforces.
Three states. Three models. A California employer faces a different set of obligations than a Colorado employer, who faces a different set than an Illinois employer, and that’s before New York, Texas, and the fifteen other states with active AI legislation are added to the picture.
This is the structural consequence of the federal cancellation. A voluntary federal pre-release review framework wouldn’t have preempted state laws, it carried no preemption authority. But federal action creates a gravitational pull on state legislative agendas. Without a federal framework establishing even a soft floor, state legislators face no coordination pressure. The field is open.
The Compliance Consequence
The absence of a federal framework clarifies one thing for enterprise compliance teams: the federal floor doesn’t exist. Planning assumptions built around a forthcoming federal AI governance baseline need to be reset.
What to Watch
What replaces it is jurisdictional mapping. Colorado’s January 1, 2027 ADMT deadline is the nearest hard compliance date for private employers. California’s EO creates no obligations today, but the dashboard and CBA analysis outputs will generate data that future California legislation will reference, compliance teams with California workforces should be documenting AI-related workforce decisions now, before that data record becomes relevant in a regulatory context.
The “reverse federalism” dynamic visible in the frontier lab posture, OpenAI lobbying state-by-state rather than at the federal level, reflects a strategic choice to engage regulation where it’s most tractable, not where it would be most comprehensive. State lobbying is more resource-intensive than federal lobbying. The fact that frontier labs are doing it anyway signals that they expect the state-level arena to matter more than the federal level for the foreseeable future.
Compliance teams should read that same signal. The federal regulatory vacuum isn’t a pause. It’s a structural condition.
What Comes Next
Three federal levers remain active. The CAISI voluntary agreements represent the current administration’s preferred governance instrument, cooperative, non-binding, and structurally unable to create the compliance pressure that even a voluntary EO review window would have imposed. Congressional AI legislation proposals exist, including the Blackburn bill and related efforts, but no federal AI legislation has achieved floor consideration in this session. The timeline for any enacted federal framework runs through the congressional calendar, realistically Q3 or Q4 2026 at the earliest, and that assumes legislative priority that hasn’t materialized.
The compliance posture that matches this landscape is state-first. Colorado’s January 1, 2027 deadline is real and imminent. California’s policy infrastructure will produce data and analysis that feeds the next legislative session. Illinois’s eight-bill package will reach some resolution in the summer session. These are the triggers to track.
The federal framework that died on May 21 was the lightest possible version of federal AI governance. Its cancellation doesn’t mean the U.S. has rejected AI regulation, it means the U.S. has declined to set a floor. States are setting their own floors, at different heights, with different instruments. That’s the compliance landscape for 2026 and into 2027.