The Center for Advancing Innovation and Standards for Super Intelligence has an annual budget of about $10 million, a staff small enough to fit in a single conference room, and a mandate to assess whether the most powerful software ever built is safe to release. It also has no permanent director. Its research reports stopped appearing in public this year, and according to two former government officials who spoke to the Associated Press, its staff have recently been asked to report informally to an office of the White House.
That last detail carries the weight, because reporting lines are how independence is either preserved or quietly lost. CAISSI sits inside the Commerce Department as an independent division. The White House is the institution whose political interests the center is meant to be able to contradict. An informal arrangement is softer than a reorganization and harder to challenge, because no formal order exists to be reviewed, published or contested.
The vacancy is not an accident of hiring. In April the division’s director was appointed and then dismissed within days. A replacement lasted a few months before resigning. The head of the parent institute now doubles as acting director, which is the arrangement an organization adopts when it cannot fill a job. In August, Congress began asking why the center had stopped publishing its reports on how safely certain AI systems performed, and the answer it received did not settle the question.
The White House’s response to questions from the AP did not address the reporting arrangement. It said the president had formally established a Super Intelligence Force, that a report on risks had been commissioned from the President’s Council of Advisors on Science and Technology, and that the administration would ensure continued American leadership. The framing is deliberate. Safety is presented as a national security function, handled by the executive, described in the language of advantage rather than assurance.
There is an argument for proximity. A testing body with a $10 million budget is roughly one-eighth the size of its British counterpart, and a small office far from the West Wing does not win budget fights. Officials who favor the arrangement say that sitting closer to the president gives the work a prominence and a funding stream it would not otherwise have, and that a finding read by the national security adviser has more effect than a finding posted on a website.
The counterargument is about what happens when a result is inconvenient. Evaluation work produces conclusions that embarrass the companies being evaluated and, by extension, an administration that has bet on those companies. A center reporting to the White House is a center whose conclusions pass through the people with the most to lose from them. The specific fear voiced by critics is not that reports will be falsified. It is that they will be delayed, softened or simply not published, and that the absence will be indistinguishable from an absence of findings.
The comparison that officials in the field keep making is to crash testing. The National Highway Traffic Safety Administration buys cars, runs them into walls, publishes the results and writes standards from them, and no manufacturer is permitted to conduct that test on its own behalf and call it compliance. AI evaluation has no equivalent, partly because the object being tested changes after it ships and partly because the tests themselves are contested. The result is a system in which the safety case for a model is assembled by the company that sells the model, and reviewed, if at all, by an agency with a staff smaller than the model’s documentation team.
The vacuum has been filled unevenly by the private sector. In September the president asked leading AI companies to undertake what he called tremendous self-policing under a voluntary pact, without saying how compliance would be monitored. Anthropic and OpenAI have both pointed to outside evaluators such as METR as evidence of seriousness, and those firms do real work. But the companies choose which evaluators to hire, what access to grant and whether to release the results, which makes an outside audit a form of disclosure rather than a form of regulation.
That asymmetry is the substance of the story. A government that cannot test models has to accept the industry’s account of how they behave, at the moment when agents are being deployed to act without supervision and incidents are being reported that would have been hard to imagine two years ago. The center’s budget, its vacancies and its reporting line are all downstream of one decision: whether the United States intends to maintain an independent evaluator at all, or whether evaluation is a service that vendors buy and publish when it suits them. Publication of the risk report commissioned from the science advisers will be the visible event, and it will be read closely by people looking for a signal about how the next round of model releases will be treated. The more informative signal is quieter. Watch whether CAISSI hires a director, whether it resumes publishing its evaluations, and whether the next finding that makes someone uncomfortable appears in public at all. A testing office is judged by what it finds. A testing office with no findings, and no one to sign them, is a letterhead.
