A US government centre responsible for testing advanced artificial intelligence systems is operating without a permanent director and has reportedly stopped publishing safety research. The developments are raising concerns about whether scientific oversight could become more closely aligned with White House priorities and the interests of the technology industry.
The situation has wider relevance for Europe because the European Union is taking a different approach through its AI Office, a body within the European Commission with enforcement powers under the EU AI Act. The contrast highlights competing models of AI governance: voluntary industry commitments in the United States and legally backed supervision in the EU.
What is happening at the US AI safety centre?
The Center for Advancing Innovation and Standards for Super Intelligence, known as CAISSI, is part of the US Department of Commerce. According to reporting by the Associated Press, the centre has no permanent leader and has stopped releasing some of its research on advanced AI risks.
Two former government officials told the news agency that staff had recently been asked to report informally to a White House office. The arrangement has not been publicly explained in detail, and the White House did not respond to questions about the centre’s management.
CAISSI’s work covers risks that could arise from highly capable AI systems, including cyberattacks and biological threats. Its research is intended to provide technical evidence that can inform public policy and safety standards.
Leadership changes have added to uncertainty
The centre has experienced rapid turnover at the top. Collin Burns, a former Anthropic researcher, was removed as director after four days in April, reportedly because of concerns about his links to the company. His replacement, Chris Fall, resigned three months later.
Arvind Raman, who leads CAISSI’s parent organisation, the National Institute of Standards and Technology, is now acting director. In August, members of Congress questioned why the centre had stopped publishing safety reports.
The leadership gap matters because an independent director can help protect technical assessments from political or commercial pressure. Without a permanent head and a clear public reporting structure, outside observers may find it harder to assess the centre’s priorities and conclusions.
Why critics are concerned about AI oversight
Critics fear that closer links to the White House could politicise the centre’s work or soften findings that are inconvenient for AI developers. The concern is particularly significant because the administration has encouraged companies to rely heavily on voluntary self-policing.
In September, President Donald Trump urged AI companies to commit to “tremendous self-policing” under a voluntary agreement. The administration has not publicly detailed how compliance with that approach would be independently monitored.
Amos Toh of the Brennan Center for Justice told the Associated Press that treating AI safety primarily as a national security issue could push other public interests into the background, including:
- Consumer protection
- Privacy and data rights
- Civil rights
- Public accountability
Caroline Siegel Singh of the Federation of American Scientists argued that policy should be shaped by nonpartisan or bipartisan technical experts acting in the public interest. The disagreement reflects a broader debate over whether AI oversight should be led mainly by national-security agencies, economic regulators or independent scientific bodies.
How the US approach compares with the EU AI Office
The EU has established a more formal regulatory structure through the European Commission’s AI Office. The office is responsible for supporting implementation of the EU AI Act and supervising certain obligations for the most advanced general-purpose AI models.
The office has had powers to fine developers of the most advanced systems since August, according to the source report. It employs more than 125 people and was reportedly recruiting an additional 40 staff. CAISSI, by comparison, receives approximately $10 million (€8.94 million) annually—around one-eighth of the budget of its UK counterpart.
The comparison should not be read as evidence that one system has resolved every challenge. The EU framework also depends on detailed implementation, cooperation with national authorities and the ability to apply rules consistently. However, the institutional difference is clear:
- US model: greater reliance on voluntary commitments and shifting executive oversight.
- EU model: statutory obligations, Commission supervision and potential penalties for covered violations.
- Shared challenge: ensuring that technical assessments remain credible, transparent and sufficiently independent.
What this means for European AI policy
The developments in Washington are likely to be watched closely by European policymakers, companies and researchers. The EU is attempting to position itself as a regulator of advanced AI while also seeking investment and innovation in the sector.
If US safety research becomes less transparent, European regulators could face greater difficulty comparing technical assessments across jurisdictions. This may affect cooperation on model testing, cybersecurity, biosecurity and standards for high-risk systems.
Companies operating in both markets may also encounter different compliance expectations. Developers could face legally enforceable requirements in the EU even where similar obligations remain voluntary in the United States. That makes regulatory clarity and cross-border coordination increasingly important.
The issue is relevant to Ireland because the country hosts major technology operations and is subject to EU digital regulation. Irish-based companies that develop, deploy or distribute covered AI systems may need to follow the EU AI Act regardless of how US oversight develops. The practical effect will depend on the company’s role, the type of system involved and the applicable provisions of the legislation.
What happens next?
The immediate questions concern CAISSI’s leadership, its reporting arrangements and whether its safety research will resume. Congress has already sought explanations about the publication gap, while the White House has pointed to a separate Super Intelligence Force led by National Intelligence Director Jay Clayton.
Further clarity is also needed on how the administration intends to monitor voluntary commitments by AI companies. Without transparent evaluation and public reporting, it may be difficult to determine whether self-policing arrangements are effective.
In Europe, attention will remain focused on the implementation and enforcement of the EU AI Act. The EU AI Office’s resources, independence and ability to coordinate with national authorities will be central to whether the bloc’s rules produce consistent oversight.
Conclusion
The uncertainty surrounding the US AI safety centre illustrates why institutional independence and public reporting matter in AI regulation. As Washington considers a closer relationship between safety oversight and the White House, the European Union is relying on a formal regulatory framework with enforcement powers. For European businesses and policymakers, the central question is whether international cooperation can preserve credible, transparent AI safety assessments while protecting consumers, privacy and civil rights.