What exactly does the executive order create?

According to the New York Times, the order establishes an AI Safety Institute (AISI) within the Department of Commerce that will review frontier AI models before public release. The institute can issue "safety determinations" that carry legal weight, though the exact enforcement mechanism remains vague. Reuters reported that the order also mandates a new reporting requirement for any model trained using more than 10^26 FLOPs — a threshold that captures all major frontier models from OpenAI, Anthropic, Google, and Meta.

The order does not ban any technology outright. Instead, it creates a review process that the AISI must complete within 90 days of a company submitting a model for evaluation. If the institute finds unacceptable risks — for example, autonomous capabilities or dual-use potential — it can recommend the model not be released until mitigations are in place.

Why did the White House abandon its hands-off approach?

The shift is dramatic. For most of his first term and the first year of his second term, Trump's AI policy was defined by deregulation — he revoked Biden's 2023 AI executive order on day one and repeatedly said government should not stand in the way of American innovation. According to the New York Times, what changed was a series of private briefings from intelligence officials who warned that unmonitored frontier AI models could be used for cyberattacks, bioweapon design, or autonomous warfare. The final straw, sources told Reuters, was a classified assessment that a major frontier lab had accidentally created a model capable of basic autonomous hacking — the lab refused to disclose details publicly, but the incident convinced national security advisors that self-regulation was insufficient.

Industry reaction has been split. OpenAI publicly welcomed "clarity" but privately lobbied for narrow review criteria, while Anthropic's CEO Dario Amodei said the order was "a reasonable step" but warned that underfunding the AISI would turn it into a bottleneck. The American AI Innovation Alliance, a trade group representing Google, Microsoft, and Meta, called the order "premature" and asked for a six-month comment period.

Who wins and who loses under this order?

This table is based on reporting from the New York Times and Reuters as of June 2–3, 2026.

How does this compare to the EU AI Act?

The EU AI Act, which took effect in 2025, takes a risk-category approach: unacceptable risk (banned), high risk (conformity assessments), and limited risk (transparency). Trump's order is more targeted — it only applies to models above the FLOP threshold and does not create a general framework for all AI applications. According to Reuters, the White House explicitly rejected a European-style tiered system, calling it "too bureaucratic." Instead, the U.S. model is essentially a frontier-model review gate. This means the U.S. is now regulating AI more than it did a year ago, but less comprehensively than the EU. The practical effect: companies building frontier models face a U.S. review process that could take up to 90 days per model, while EU rules require ongoing compliance across all product tiers.

What remains uncertain about enforcement?

The order's text is notably silent on penalties for non-compliance. According to the New York Times, the AISI can "recommend" that a model not be released, but it is unclear what happens if a company ignores that recommendation. The order directs the Commerce Department to propose enforcement rules within 120 days, but until those rules exist, the AISI's authority is largely moral suasion. Another uncertainty: the order applies to models trained after the effective date, but what about models already in development? Reuters reported that several labs have asked for grandfather clauses, which the White House has not addressed. Finally, the order does not cover open-source models that are fine-tuned after release — a massive loophole that Meta and others could exploit.

1. By Q1 2027, the AI Safety Institute will have reviewed exactly zero models — the 90-day review clock will not start until enforcement rules are finalized, which will slip past the 120-day deadline.
2. OpenAI will delay the public release of GPT-6 by at least six months, citing the need to align with new regulatory expectations, effectively giving Anthropic a window to capture enterprise market share.
3. Meta will announce that future Llama models will be released under a "research-only" license to avoid AISI review, a move that will draw criticism from both safety advocates and open-source purists.

• The order creates a review gate for frontier models, but enforcement rules are not yet written — the real test comes in the next 120 days.
• Anthropic is positioned as the regulatory winner because its existing safety processes align with the new requirements; Meta and OpenAI face the most disruption.
• The order does not address open-source fine-tuning, a loophole that could render frontier-model review ineffective if labs simply release base models and let the community fine-tune them.
• China will use this order as propaganda to argue that U.S. regulation proves Beijing's state-led approach is faster, even if the evidence for that claim is weak.
• Legal challenges are almost certain: the order's statutory basis is thin, and at least one major lab will test it in court by year-end.