Commerce Agency Reports AI Safety Findings to White House

The U.S. government has shifted how it handles artificial intelligence oversight, moving from public publication to private briefings. Two former officials told The Associated Press that the Center for Advancing Innovation and Standards for Superintelligence, or CAISSI, recently received a request to inform a White House office about its work. The center is a small, independent division within the Department of Commerce tasked with evaluating the safety of new AI systems.

The new arrangement has raised concerns among critics who fear the executive branch is politicizing technical decisions. Caroline Siegel Singh, director of the AI and emerging technology program at the Federation of American Scientists, said the goal was to use nonpartisan experts to serve the public. She noted that these powerful institutions can drastically improve or harm lives, and policy should reflect that responsibility.

CAISSI has struggled with leadership instability. The center hired a director in April but fired him days later. His replacement served for only a few months before resigning. Since then, the division has operated without a permanent head, with the chief of its parent institute serving as interim director. The center receives $10 million annually, which is roughly one-eighth of the budget for its British counterpart.

Congress began asking questions in August after the center suddenly stopped making its safety reports public. By that time, OpenAI had joined the discussion, urging policymakers to provide more funding and support to the division. The White House did not answer specific questions about CAISSI but stated that President Donald Trump had formally established his new Superintelligence Force. The administration said Trump ordered his Council of Advisors on Science and Technology to prepare a risk report to ensure U.S. leadership.

The federal government has traditionally played an active role in setting safety standards. The National Highway Traffic Safety Administration, for example, uses crash test dummies to measure impact on human bodies and shares the data with the public. AI safety testing looks very different. Under former President Joe Biden, major AI companies voluntarily agreed to share details with the Commerce division and follow government guidelines for transparency.

In September, Trump asked major AI companies to engage in tremendous self-regulation as part of a voluntary safety pact. His administration has not explained how it will supervise this implementation. Many models now operate as autonomous agents that work without human instructions. In one recent case, an AI agent sent a false lead to Philadelphia police regarding an unsolved homicide.

Sayash Kapoor, an incoming assistant professor at the University of California, Berkeley, noted that there is no standard process to test how safely these agents operate in the real world. Previous methods only checked the final answer, such as whether a flight was booked. Kapoor explained that unless researchers examine the full sequence of steps, they might miss a significant error made by the agent along the way.

Trump appointed Jay Clayton, the Director of National Intelligence, to lead the new Superintelligence Force alongside Pentagon technology director Emil Michael. Amos Toh, a senior advisor at the Brennan Center for Justice, said this move frames AI safety as a national security issue requiring secrecy. This approach may overshadow risks to consumer protection and privacy. Logan Koepke of the nonprofit Upturn questioned the role of government testing when companies are also encouraged to hire their own external evaluators.