Staff at the federal center that tests the safety of advanced artificial intelligence systems were recently asked to report informally to a White House office, The Associated Press reported Sunday, citing two former government officials. The arrangement brings the government's main independent AI model evaluator closer to political control while it lacks a permanent leader and has stopped publishing research, according to AP.
What Happened
The unit is the Center for Advancing Innovation and Standards for Super Intelligence, or CAISSI, part of the Commerce Department's National Institute of Standards and Technology. Until recently it was called the Center for AI Standards and Innovation. The two former officials spoke to AP on condition of anonymity to avoid repercussions for current staff.
AP said the center receives about $10 million a year, roughly one-eighth of what its British counterpart gets, and has lost staff. Its director was hired and fired within days in April, and his replacement resigned after a few months. Since then, the head of its parent institute has also served as acting director. In August, Congress asked why the center had stopped publishing reports on how safely certain AI systems performed, AP reported. The center's website still lists a Sept. 17 assessment of the cyber capabilities of a model from Chinese developer Z.ai.
The White House did not answer AP's specific questions about the center. In a statement, it said President Donald Trump had formally established his Super Intelligence Force and that his Council of Advisors on Science and Technology was preparing a report on AI risks.
Why It Matters
Government testing is one of the few outside checks on frontier AI models, the most capable general-purpose systems. CAISSI describes itself as industry's primary point of contact in the U.S. government, with evaluations centered on cybersecurity, biosecurity and chemical weapons risks. Under the Biden administration, leading AI developers voluntarily agreed to share model details with the center and follow its testing guidelines.
Critics told AP they fear the executive branch could politicize findings on issues such as cybersecurity and bioweapons, or soften them in deference to the industry. Amos Toh, senior counsel at the Brennan Center for Justice, told AP such decisions "should be explained and rationalized in the public eye." Others argue that closer ties to the White House could raise the center's profile and funding.
For boards, the question is whether the evaluations they cite when approving AI deployments stay independent and public. Testing that answers to political officials could also change what developers are willing to share with evaluators. In September, Trump asked AI companies for "tremendous self-policing" under a voluntary safety pact, and the administration has not said how it will check compliance, AP reported.
Context
The Primary Wire reported Oct. 12, citing Reuters and The Information, that the White House planned to move three of the center's officials to report to science adviser Michael Kratsios and National Cyber Director Sean Cairncross. Sens. Maria Cantwell and Mike Rounds later urged more resources and a broader mandate for it. Toh told AP that Trump's choice of National Intelligence Director Jay Clayton to lead the Super Intelligence Force shows the administration framing model safety as a national security matter.
AP also noted there is still no standard process for testing how AI agents, systems that carry out tasks on their own, behave in the real world. Sayash Kapoor of UC Berkeley, co-author of a new paper on agent evaluations, told AP that checking only final results can miss serious errors.
What to Watch
Watch whether the informal reporting line becomes formal, whether a permanent director is named and whether the center resumes publishing evaluations. Incoming Democratic leaders plan bills on independent testing of advanced models, The Primary Wire reported Nov. 6. Any such bill would have to settle who does that testing and how insulated it is from the White House.
