U.S. Government Urges Meta to Accept AI Safety Review, Making It the Only Major Developer Without an Agreement
Alina Collins
The Trump administration is pressuring Meta to voluntarily submit its AI models for federal safety review; Meta is the only one of six major US developers that has not signed an agreement, putting its open-source AI strategy under direct pressure from Washington.
What exactly is the government asking Meta to do?
The Center for AI Standards and Innovation (CAISI), housed under the Commerce Department, has emailed Meta requesting it voluntarily submit AI models for safety evaluation.
This means → the government wants to see and test models before they go public, probing for capability limits and security flaws.
OpenAI, Anthropic, Google, xAI, and Microsoft have all signed voluntary sharing agreements — Meta is the sole holdout.
Why hasn't Meta signed?
A Meta spokesperson declined to comment immediately; sources say Meta's policy team is negotiating terms with the Commerce Department, and whether a deal will materialize remains unclear.
In plain terms = Meta released its latest model Muse Spark in April, performing near the level of Google and OpenAI peers — the issue is not unwillingness to cooperate, but how to cooperate.
This reflects a deeper tension: Meta's open-source approach means anyone can use its models once released. A pre-release government review clashes with the speed and openness that define that strategy.
What did the executive order change?
On June 2, Trump signed an executive order requiring tech companies to give the government up to 30 days to evaluate models before public release, with a full review process to be established by the end of July.
Key details remain undefined: who leads the review and what standards models will be judged against are both unresolved.
This means → the executive order drew a frame, but the rules inside it are still blank — companies face an exam sheet that hasn't been finished yet.
Why is Anthropic's experience a warning?
Anthropic submitted its flagship model Fable 5 for government review, yet last week the White House gave it fewer than 90 minutes to shut down model access — after Amazon researchers published a paper exposing a vulnerability that could enable cyberattacks.
In plain terms = even if you cooperate fully, the government can still demand you pull the plug on very short notice — cooperation does not equal safety.
Trump told Axios over the weekend he no longer views Anthropic as a security threat, and talks are progressing; but who ultimately approves restoring access, and whether other companies will face the same standard, remain unclear.
What does this mean for the industry?
The "voluntary" mechanism is being redefined by the executive order and the Anthropic incident: shifting from a goodwill gesture to a gatekeeping threshold with real enforcement.
Whether Meta can find acceptable agreement terms within this framework is the next critical test case.
This means → for every AI developer, the question is no longer "should we cooperate with review" but "once we cooperate, where does the government's authority end?"
Content is for reference only, not financial advice.