WASHINGTON — The White House has finalized a new AI safety framework but plans to keep its evaluation criteria for advanced AI models private. Only a select group of large AI companies will receive the detailed standards, allowing them to adapt their operations accordingly. The decision has drawn criticism for creating an uneven playing field in the artificial intelligence sector.
The framework includes a 30-day pre-release review process for frontier AI models. During that period, the specific benchmarks and thresholds used for evaluation will not be made public. Researchers, smaller startups, open-weight developers and many businesses will not know the criteria by which these systems are judged.
Companies informed of the new standards include Meta, Nvidia, Microsoft, OpenAI and Anthropic. That selective disclosure gives those firms an early look at the government's expectations — which critics call a cheat sheet — while other developers operate without clear guidance.
The administration says its goal is to protect national security and strengthen the U.S. position in the AI race against China. Critics argue the approach pairs unclear governance with perceived favoritism, potentially harming American innovation.
Industry observers contend that national security and transparency are not mutually exclusive. Without public understanding of the framework's principles, they say, the broader AI community and allied nations will struggle to align with the safety measures.
Smaller AI labs and third-party researchers are particularly affected. Without access to the evaluation criteria, they face significant hurdles in developing and testing models to meet federal standards — a dynamic critics call an entrenchment program for established AI providers.
One person familiar with the White House's discussions with AI labs said the secretive process advantages larger companies and risks stifling competition by making it harder for new entrants to innovate.
The secrecy also erodes public trust. Safety frameworks depend on verification, which is difficult when the tests themselves are hidden. The arrangement creates the perception of an insider's club, particularly given that the largest AI companies may have helped shape the framework.
Applying the framework universally and giving all U.S. companies the same information would ensure a level playing field. As long as the administration keeps the rules private while granting select companies inside access, both transparency and competition are at risk.
