The U.S. government division responsible for assessing the safety of new artificial intelligence systems is facing questions about its independence. The Associated Press reported that staff at the Center for Advancing Innovation and Standards for Super Intelligence, or CAISSI, were recently asked to informally report to a White House office. The center is part of the Commerce Department, but it lacks a permanent leader and has stopped publishing its research. Critics worry that a body meant to provide technical checks could become more vulnerable to political pressure.
CAISSI was created to help evaluate powerful AI systems as companies deploy models that can plan and act with less human prompting. The AP report said the center has experienced leadership turnover, including a director who was hired and then dismissed within days and a replacement who later resigned. Its annual budget is about $10 million. Those constraints matter because independent testing requires stable staff, clear methods and enough resources to examine systems whose risks can involve cybersecurity, privacy, civil rights and even bioweapons.
The dispute reflects a wider argument about whether AI companies can be trusted to police themselves. Leading laboratories have signed voluntary safety commitments and promoted internal or external evaluations, but companies can limit access to systems and choose which findings to publish. Government testing can provide a separate public-interest check, similar in principle to how regulators test vehicles or inspect other safety-critical products. Yet that role only works when evaluators can publish evidence and explain their standards without needing approval from the organizations they are assessing.
The issue is relevant beyond Washington because AI models and agents are used across borders, including by Philippine businesses, schools and government offices. A credible safety system needs technical expertise, transparent procedures and protection from retaliation when results are uncomfortable. It also needs to keep pace with agents that perform multi-step tasks rather than only generating a final answer. The CAISSI questions show that AI governance is not only about writing rules. It is also about preserving the independence and public visibility of the people who test whether those rules work.
