Trump Delays AI Security Review Requirements Before Model Launch

On May 21st, the Trump administration decided not to sign an executive order that would have required AI models to undergo government safety reviews before public release. According to TechCrunch, the president himself justified the decision with a statement that says it all: "I don't want to get in the way of that leading", referring to U.S. technological leadership. The problem, according to the White House, lay in the specific language of the text.

It is not a definitive refusal, but a postponement. Yet in an industry where development cycles are measured in weeks, the difference between signing today or three months from now could mean several models launched without the oversight framework the draft intended to establish.

What the Order Would Have Required

The delayed draft would have introduced a pre-launch review requirement by government agencies competent in national security and critical infrastructure matters. The specific mechanism has not been fully made public, but the logic was clear: before a frontier model reached the market, the government would have a window to assess systemic risks.

This kind of pre-deployment oversight already exists in various forms for pharmaceuticals, autonomous vehicles, or weapons systems. Applying it to generative AI models was, for its supporters, a logical step given the pace of adoption. For its detractors, among them apparently sectors within the presidential circle itself, it was bureaucratic red tape that could slow U.S. companies just as China accelerates.

Why It Matters and for Whom

For safety teams at companies like Anthropic, OpenAI, or Google DeepMind, the order would have created a formal interlocutor in Washington with authority to flag vulnerabilities before launch. That has an upside—institutional legitimacy for internal safety evaluations—and a downside: a regulatory process that could have become a lever for political or competitive pressure.

For engineering teams integrating these models into products, the typical ClaudeWave reader, the most direct consequence would have been uncertainty around availability dates for new versions. An undefined government review process is, in practice, a planning risk.

The postponement, paradoxically, leaves things as they were: no mandatory federal framework, with the laboratories themselves managing their internal evaluation processes—red teaming, dangerous capability assessments, safety reviews—voluntarily and by their own standards.

The Broader Picture

This decision comes after the Trump administration rescinded, in early 2025, Biden's AI executive order, which included notification requirements for large-scale models. Since then, federal policy has tilted more toward competitiveness and less toward preventive oversight. The May postponement confirms that trend, though it does not entirely close the door to some form of future regulation.

In Europe, the AI Act is already in the phased implementation stage, with specific requirements for high-impact general-purpose models. The regulatory divide between the two sides of the Atlantic continues to widen, complicating matters for companies operating in both markets.

From our perspective, the takeaway is nuanced: a government wanting not to hinder innovation is understandable, but delegating risk management entirely to the laboratories themselves, without any independent verification mechanism, is a bet that requires those laboratories to maintain high standards even when there are no external consequences for lowering them. Whether that happens consistently remains to be seen.