The center that tests AI now answers to a White House office.
By Callahan Blackwood | November 8, 2026
The federal center charged with independently testing America's most powerful AI systems has stopped publishing its findings. Its officials now report informally to a White House policy office. It has had no permanent director since spring. Its annual budget is $10 million, roughly one-eighth of what Britain allocates to the same function.
That is where the Center for Advancing Innovation and Standards for Super Intelligence stands today, according to reporting by the Associated Press published Sunday, November 8, through EuropeSays and the Plain Dealer wire.
The center is known as CAISSI. It operates inside the National Institute of Standards and Technology, a division of the Commerce Department. Its stated mandate has not changed: assess whether new AI systems are safe before those systems reach broad deployment. The operational picture surrounding that mandate has changed, and those changes raise a structural question that the administration has not publicly answered. If the office that tests AI answers to the same White House apparatus that promotes AI, who independently validates the tests?
A Center Remade Several Times Over
CAISSI is not a new idea, and it has not had a stable identity.
It began as the AI Safety Institute under the Biden administration. It was then renamed the Center for AI Standards and Innovation. It was renamed again to its current form: the Center for Advancing Innovation and Standards for Super Intelligence. Each rebrand carried a policy signal, a shift in emphasis from safety caution toward innovation framing and competitive positioning. The sequence of names tracks an ongoing argument within the federal government about the center's primary purpose.
That history matters because the name changes were not cosmetic. Each transition reshuffled leadership priorities, organizational reporting lines, and, eventually, publishing habits. What the center produces, and who sees it first, has changed along with its name.
The most visible change is the halt in publications. The AP, citing government officials, reported that CAISSI's research output has gone dark. No new assessments of AI systems have been released publicly. The halt is documented. The stated rationale is not.
The Leadership Churn
The timeline of turnover at CAISSI is short and consequential.
A director was hired in April 2026. He was fired within days of taking the role. His replacement took over and served for several months before resigning. The center currently has no permanent director. The acting director holds that position concurrently with leading NIST's parent institute, meaning the person nominally running CAISSI carries a separate, broader set of institutional responsibilities at the same time.
That sequence plays out across roughly seven months. For any technical body, sustained leadership is not a bureaucratic nicety. For a center evaluating frontier AI systems on behalf of the federal government, it is a functional requirement.
Assessing whether a large-scale AI system behaves as its developers claim is not a task that resets cleanly with each new director. It requires continuity. Evaluators need institutional memory of prior assessments, comparative context from earlier models, and established working relationships with the developers submitting systems for review. Those relationships take time to build. They also depend on the center having a credible track record, which is itself built through consistent, published output over time. Each departure disrupts that accumulation. Two departures in seven months, with no permanent replacement in place, interrupts it substantially.
Congress noticed. In August 2026, legislators formally questioned why CAISSI had stopped publishing its research. That inquiry is part of the documented record. The White House did not answer the AP's specific questions about the publication halt or the reporting structure change. Administration officials pointed instead to the president's Super Intelligence Force and a commissioned risk report. Those references did not address the questions asked.
Answering to a White House Office
The informal reporting directive is the structural fact at the center of the accountability question.
CAISSI officials have been asked to report informally to a White House office. That is not a formal reorganization on the administrative record. It is an informal directive, according to the AP's reporting. The distinction has practical consequences. A formal reorganization is visible, documented, subject to administrative procedure, and potentially subject to congressional oversight. An informal reporting arrangement is harder to track, and harder for oversight bodies to scrutinize.
The White House office in question sits within the broader apparatus the administration has built around AI governance. Jay Clayton, the National Intelligence Director, leads the Super Intelligence Force, which is the administration's primary vehicle for AI policy. When the AP asked about CAISSI's situation, administration officials pointed to the Super Intelligence Force and to a commissioned risk report as evidence of the administration's engagement with AI risk.
That response is a non-answer to the specific questions reporters asked. It does not explain why the center stopped publishing research. It does not explain why two directors departed in rapid succession. It does not describe what the informal reporting relationship means for the center's independence from the policy decisions its assessments are meant to inform.
What an AI Evaluation Center Does, and Why Independence Is the Point
To understand what is at stake, it helps to be specific about the work CAISSI is designed to do.
When a frontier AI system is submitted for assessment, evaluators test it across a range of risk dimensions. They probe whether the system can be used to generate knowledge useful for producing weapons. They test whether it behaves consistently when subjected to adversarial prompting designed to surface unintended capabilities. They compare the system's observed behavior against the claims its developers have made about its safety properties. Those tests are technical, time-consuming, and require direct access to the model, including in some cases access to its underlying weights and architecture.
The findings from those assessments feed into policy decisions. They inform judgments about whether a system should be cleared for broad deployment, whether it warrants restrictions, and whether a developer's public safety representations hold up under independent scrutiny. That function only works if the people conducting the evaluations operate at arm's length from the people who set policy based on the results.
That is the independence problem stated plainly. If CAISSI reports informally to a White House office that is simultaneously managing relationships with the AI industry and advocating a "tremendous self-policing" approach to AI safety, the evaluators are embedded in the policy apparatus. That does not mean findings are changed. But it removes one layer of structural insulation between the people doing the testing and the people with interests in the results. Independence is what makes an assessment credible to anyone who did not conduct it.
The budget figure sits alongside the structural concern. CAISSI operates on $10 million per year. Its British counterpart operates on roughly eight times that amount to perform a comparable function. The gap limits the number of models that can be evaluated in any given period, the depth of each evaluation, and the center's ability to retain specialized staff in a labor market where advanced AI expertise commands significant compensation. A center running on $10 million annually, with no permanent director and a halted publication record, faces serious constraints on the scope and credibility of its evaluation program at the moment when frontier AI systems are being deployed at scale.
The September Voluntary Pact
One month before this reporting, the administration called on AI companies to undertake what it described as "tremendous self-policing." Companies were invited into a voluntary safety pact. No public account of how compliance with that pact is monitored has been released.
A voluntary pact monitored by its signatories is not an independent accountability mechanism. External verification gives such agreements meaning: an independent body capable of checking whether the commitments made are being honored, and capable of reporting publicly when they are not. CAISSI, in its prior form, was a natural institutional candidate to perform or at least inform that verification function. Whether it is in a position to do that now, given its operational state, is a question the administration has not addressed.
The juxtaposition is uncomfortable. In September, the administration called for industry to police itself. In the same period, the government body best positioned to provide external oversight of that self-policing had halted its publications and been directed to report to a White House office.
Questions the Reporting Leaves Open
The AP's reporting establishes a documented set of facts. It does not resolve every question those facts raise. Some of those questions deserve to sit on the record as questions.
Has any AI safety assessment been changed, softened, or buried as a result of the new reporting structure? Nothing in the reporting establishes that any finding was changed or buried. That statement is explicit, and it matters. The concern the reporting raises is structural, not evidentiary. The structure of accountability can be compromised before any specific finding is affected, and that structural compromise is worth examining on its own terms.
Who authorized the informal reporting directive, and when? A formal reorganization generates documentation that oversight bodies can request. An informal directive may not. If members of Congress were already asking about the publication halt in August, the provenance of the reporting change is a natural follow-on question.
What is the operational relationship between the Super Intelligence Force and CAISSI's evaluation function? The administration pointed to the Super Intelligence Force in response to direct questions about the center. If that body now sets direction for CAISSI's work, the evaluations that come out of CAISSI will carry that association. How the Super Intelligence Force and the evaluation center divide their responsibilities is not publicly described.
Is the September voluntary safety pact being monitored by any external body? If not, what is the enforcement or verification mechanism? The administration has not answered this in public.
None of these questions are answered in current public reporting. They are worth asking plainly.
The Counter-Argument
Not everyone who follows federal AI policy reads CAISSI's current situation as straightforwardly damaging.
Some observers argue that proximity to the White House could benefit the center. A federal office that feeds its assessments directly to senior policymakers carries more weight in inter-agency disputes over resources and regulatory jurisdiction than one whose reports surface in a public archive. On that view, an informal reporting relationship to the White House is a channel for influence, not a threat to independence. If CAISSI's findings reach the administration directly, they might shape decisions more reliably than published reports that go unread.
There is also a pragmatic case about funding. The $10 million budget is a real constraint. Critics of the current arrangement point to it as evidence of institutional neglect. But if White House proximity translates into stronger advocacy for a larger appropriation in future budget cycles, the center's operational capacity could improve significantly. A better-funded center that works within the administration's orbit may ultimately evaluate more AI systems than an underfunded one with cleaner formal independence.
These arguments are worth taking seriously. They do not dissolve the structural question, but they complicate the assumption that every organizational change at CAISSI represents straightforward deterioration. The tradeoffs are real, and they have not been adjudicated publicly.
What Has Been Established
The Center for Advancing Innovation and Standards for Super Intelligence has undergone two name changes, two director departures in seven months, a halt to public research output, an informal directive to report to a White House office, and a congressional inquiry, all while operating on a budget that is one-eighth of its British counterpart.
The administration's response to press questions was to reference the Super Intelligence Force and a commissioned risk report, neither of which addresses what was asked.
Nothing in the reporting establishes that any specific AI assessment was altered, suppressed, or softened as a result of these changes. That is an important limit on what can be claimed.
What can be said is structural. Independent technical assessment of AI systems depends on evaluators who are insulated from the policy decisions their findings inform. That insulation is built through formal organizational independence, sustained leadership, a consistent publication record, and adequate resources. The documented record on all four counts at CAISSI gives reason for scrutiny. The administration has not provided a public account of how that scrutiny should be satisfied.
The voluntary safety pact signed in September asks AI companies to police themselves. The question of who polices the pact remains open.
