America Has Two AI Governance Regimes. They Contradicted Each Other Within Ten Days.

On June 2, 2026, President Trump signed an executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security." The order established a voluntary 30-day pre-release review window during which AI developers may submit frontier models for government cybersecurity evaluation. It explicitly barred mandatory licensing. It created an AI cybersecurity clearinghouse. It was praised by every major AI lab in the country.

Anthropic's statement called it "an important step in strengthening America's leadership in AI." Anthropic had already been running a version of this cooperation model through Project Glasswing — advance access to Mythos 5 for government-aligned partners before the EO existed.

Ten days later, the Commerce Department issued an export control directive suspending Claude Fable 5 and Claude Mythos 5 globally for all customers, everywhere, immediately. No voluntary framework. No 30-day review window. No appeal process described in the public record. The most cooperative major AI lab in the US government relationship had its flagship models taken offline.

These two actions came from the same administration. They are not the same policy.

What the EO actually says

The EO's innovation framework is real. The "frontier model" voluntary review creates a 30-day window before public release during which labs may submit models for government cybersecurity evaluation. The framework has to be developed by Treasury, NSA, and CISA by August 1, 2026 — it does not yet exist as operational procedure as of today (June 24). The "frontier model" designation itself will be determined by a classified benchmark not subject to public comment or challenge.

The no-mandatory-licensing clause is explicit: "it does not create any mandatory governmental licensing, pre-clearance, or permitting requirement for the development, publication, release, or distribution of new AI models." This is a deliberately strong statement. The White House, through OSTP, was sending a signal to the AI industry: the US regulatory posture is innovation-first, voluntary cooperation, no EU-style pre-market gatekeeping.

The 30-day window is itself a compromise. Earlier drafts ran 90 days. The shorter window reflects White House concern that a longer period would disadvantage US labs relative to Chinese competitors releasing without any government review. The policy logic: maintain government visibility into frontier models without creating a regulatory delay long enough to become competitively harmful.

What the export control directive actually did

The June 12 export control directive used Commerce Department authority under export control law — a completely different legal instrument from the EO — to classify certain AI model capabilities as export-controlled technology. The immediate operational consequence: Anthropic could not continue serving Fable 5 or Mythos 5 to anyone, including domestic US customers, because the compliance infrastructure to distinguish permitted from restricted users did not exist at that point in Claude's deployment stack.

The EO's voluntary framework and the export control authority are not in the same legal namespace. The EO operates under presidential authority to direct executive agencies. Export controls operate under the Export Control Reform Act, which gives Commerce independent statutory authority. The administration can pursue both simultaneously without one overriding the other — and that is exactly what happened.

The result: Anthropic, the company that had been operating under the spirit of the EO before the EO existed, faced a punitive national security action that the EO's voluntary framework offered no protection against.

The voluntary compliance problem

Every major AI lab publicly praised the EO. Anthropic, Google, OpenAI, Microsoft — unanimous public support for a government AI policy is so unusual it is worth examining what it signals.

The CFR's expert analysis identified the mechanism: labs "will likely participate if only to forestall more invasive regulation later." The voluntary framework is voluntary in the same sense that a business meeting with regulators is voluntary — technically optional, practically not. Labs that decline to cooperate with a voluntary government framework invite the kind of scrutiny that produces outcomes like mandatory frameworks. The rational response is enthusiastic public endorsement.

This means the EO's "voluntary" designation understates its de facto coercive force — while also overstating its protective value. Anthropic demonstrated maximum cooperation: Project Glasswing launched before the EO, public endorsement on June 2, active government partnership across 50+ organizations. None of it prevented the June 12 suspension.

The framework's effectiveness is entirely dependent on labs being honest about their capabilities — a dependency the CFR analysis identifies explicitly: "the government cannot assess what it cannot see, and frontier capabilities are visible only to the labs that build them." The EO creates a government visibility channel. It does not create government enforcement capacity.

Two governance regimes, operating simultaneously

The US now has at least two parallel AI governance structures with different legal bases, different authorities, and different operational logic.

The first is the EO innovation framework: voluntary, transparent, cooperative, review-before-deployment as an option rather than obligation, explicit prohibition on mandatory licensing. This is the framework the AI industry endorsed and the White House OSTP designed.

The second is the national security apparatus operating through Commerce export controls, NSA signals intelligence authority, and interagency classification mechanisms that are largely opaque to public view. This framework is not voluntary. It does not require advance notice. It does not have an appeal process that labs can activate. It operates on national security grounds that can be invoked without the kind of procedural transparency the EO framework creates.

Both frameworks are currently active. The first applies when the government wants to gather intelligence about AI capabilities. The second applies when the government wants to take action against those capabilities.

The Anthropic suspension is the clearest available example of both operating on the same target within the same month. Whether the voluntary cooperation under Project Glasswing informed the export control assessment — whether the government used information Anthropic provided in good faith to execute the suspension — is not public knowledge. The question has not been asked publicly. If the answer is yes, it reframes what "voluntary cooperation" means in practice for AI labs with national security-relevant technology.

The classified frontier model problem

The EO's "frontier model" determination — which labs are subject to the voluntary review framework — will be made through a classified benchmark not subject to public accountability. This is a significant governance gap regardless of the administration's good intentions.

Classified determinations of which companies are subject to government scrutiny, without external review, create conditions for selective or punitive application. If the benchmark is set to include only certain capability thresholds, it may exclude models that pose real risks while including models that don't. More importantly, it creates an authority structure — government-determined classification of AI model significance — that operates outside the procedural safeguards that make government regulation legitimate.

The Biden EO that Trump revoked was imperfect. But it had mandatory reporting requirements with defined compute thresholds (10²⁶ FLOPs) subject to public comment. The voluntary framework replaces a flawed-but-transparent mechanism with a flexible-but-opaque one.

The EU contrast nobody is drawing correctly

The standard framing pits the Trump EO against the EU AI Act as "voluntary vs. mandatory." This is accurate but incomplete. The EU AI Act is mandatory pre-market compliance for high-risk systems — bureaucratic, slow, and costly. The Trump EO is voluntary with no consequences for non-participation.

The Anthropic export control suspension adds a third category that neither framing captures: unilateral national security action that operates outside any formal regulatory framework and has no procedural constraints visible to the regulated entity. The EU does not have an equivalent mechanism. The EU cannot suspend a lab's flagship models globally overnight through an internal security determination.

The US approach is not simply "more flexible" than the EU approach. It is structurally different in a way that creates different kinds of unpredictability. EU labs face regulatory compliance costs and potential market-access restrictions. US labs face a combination of regulatory permissiveness and unconstrained national security authority. These produce different strategic calculations.

What happens August 1

The implementing agencies — Treasury, NSA, CISA — must deliver the voluntary framework's operational procedures by August 1, 2026. This is the first concrete milestone in converting the EO's principles into actual policy. The classified frontier model benchmark will be determined as part of this process.

If the framework is operationalized well, it creates a real government capability to identify exploitable vulnerabilities in frontier models before public deployment — a legitimate national security goal. If it is operationalized poorly or selectively, it creates a formal structure for the same kind of unilateral pressure the export control directive already demonstrated is available without any framework at all.

The export control suspension remains unresolved as of June 24. The resolution timeline, the appeal process (if any), and the precedent it sets for other frontier models are the open questions. They are also the questions the EO framework was designed to make unnecessary — and clearly did not.