All major AI research labs signed up for US AI safety review — except Meta

The U.S. Center for AI Standards and Innovation (an evaluations agency run by Commerce Secretary Howard Lutnick) has requested that Meta make its Muse Spark AI model available to the government for evaluation prior to its release. This is in accordance with an executive order issued on June 2nd that lays out a process for evaluating cutting edge AI models, allowing the government to evaluate a model within 30 days of its submission to determine its functionality and any associated risks.

What’s interesting is that Meta is the only one of the top 6 AI labs that hasn’t signed onto this agreement. Other labs including OpenAI, Anthropic, Google, xAI, and Microsoft have all given the government early access to their models for evaluation. When asked about the government’s request, a spokesperson for Meta said “we share the administration’s objective of supporting US leadership in developing safe and robust frontier AI,” but did not commit to following through with the request.

1How the new review framework works

In this model, AI labs agree to give the government access to their strongest models prior to releasing them to the public. Within 30 days, the government will evaluate the model based on its functionality, potential abuses, and any necessary restrictions or precautions. Essentially, this provides the government with a chance to inspect a model prior to its public release, analogous to how pharmaceuticals and medical devices are evaluated prior to market.

Meta has traditionally been more transparent than other AI labs, open-sourcing its LLaMA and Muse families of models rather than locking them behind paywall APIs. This could be contributing to the tension around the request. It’s difficult to reconcile a policy of open-releasing models to the public with a policy of giving the government exclusive access to models for 30 days prior to release.

2Why AI builders should pay attention

This is the first instance of a formal process for pre-releasing AI models in the U.S., and 5 of the 6 top AI labs are already participating. Once Meta (and potentially others) follow suit (and they likely will), there will be a new step in the AI development cycle: models must be reviewed by the government prior to being released to developers and consumers.

In the short term, this won’t affect anyone using AI APIs much — models will continue to be released, albeit with the government getting a chance to review them first. However, in the long term this represents a shift towards a more formalized relationship between frontier AI research and the U.S. government. It also has some similarities to how other high-consequence products are evaluated. It’s unclear whether this will increase safety, delay development, or change the types of models that are developed.

The particular model in question is Meta’s current strongest model, which includes both traditional and extended reasoning abilities (Muse Spark). The model was released in April, and the government appears to be interested in seeing how Meta’s reasoning-enabled model works prior to its broader release.