White House AI review plan covers closed models, not open ones

In short: The U.S. has finalized a voluntary review process for some powerful AI models, but it only applies to closed models and excludes models released with public code or weights.

What happened

The White House has rolled out a new voluntary “frontier AI” review framework under Executive Order 14409. It is meant for very capable AI systems that could raise national security or cyber risk.

Under the plan, developers of certain “covered frontier models” can choose to give the U.S. government up to 30 days of access before the model is released. This review is run through NIST and the Center for AI Standards and Innovation.

A key detail is who counts as “covered.” The framework defines covered frontier models as closed-source models, meaning the company does not publish the underlying code or the model’s “weights” (the stored settings that make the model work, like the full recipe behind a dish). Models that publish their code or weights are not included in this pre-release review process.

The framework also says it should not be read as restricting open models after they are released. It is not a licensing or approval system, and it does not give the government a legal veto over release.

Why it matters

This creates two tracks for powerful AI in the U.S. Closed models from major companies are the main focus of government testing before launch. Open models, even if they are very capable, can be released without going through this specific 30-day check. For the public, it affects what kinds of AI get extra scrutiny before they reach consumers, businesses, and hackers.

Source: NYTimes