US AI safety agency faces leadership crisis after director resigns
The resignation of Chris Fall as director of the Center for AI Standards and Innovation after just 90 days is not a story about a single person leaving a job. It is a story about how the U.S. government builds the infrastructure for testing the most powerful technologies ever created, and how that infrastructure is being hollowed out from the top before it can prove its worth. The AI industry does not need more benchmarks, more model cards, or more corporate safety pledges. It needs a stable, credible institution that can say ‘this model is safe enough to release’ with authority that developers and the public can trust. When the person in charge of that institution vanishes every three months, the institution itself becomes a liability rather than an asset.
The Agency That Tests the Untestable
CAISI, previously known as the U.S. AI Safety Institute, was created to solve a problem that the private sector cannot solve on its own: how to evaluate an AI system that is so complex that even its creators do not fully understand its behavior. The agency’s scientists and engineers work with companies such as OpenAI, Google DeepMind, and Anthropic to test frontier models before they reach the public They study risks that sound like science fiction but are becoming engineering reality: AI systems that could help design cyberattacks, synthesize chemical weapons, or engineer biological threats. This is not theoretical hand-wringing. These are concrete evaluation protocols that determine whether a model is released to millions of users or sent back for more work. The agency sits at the intersection of technical capability and public safety, and its director is supposed to make that intersection functional
The Revolving Door That Never Stops Spinning
Fall’s resignation marks the second leadership change at CAISI in less than a year. His predecessor, Collin Burns, left after less than a week on the job, reportedly due to tensions over his previous work at Anthropic, a company that had clashed with the administration over AI safety approaches. Now Fall is gone after 90 days, and the Commerce Department offers no explanation beyond a terse confirmation of his departure. The agency is now being run by Arvind Raman, the director of NIST and former dean of engineering at Purdue University, on an acting basis. Raman is a respected figure, but he already has a full-time job running an entire national institute. Adding CAISI leadership to his responsibilities is a temporary patch, not a solution. The department says it expects to announce a permanent replacement in the coming weeks, but the pattern suggests that the problem is not finding candidates—it is keeping them
Why Stability Matters More Than Speed
For AI companies, the uncertainty at CAISI is not an abstract governance issue but a concrete business risk. Government testing frameworks increasingly determine how quickly advanced models can reach customers. A model that passes evaluation can launch on schedule; a model that raises questions can be delayed for months while the agency decides what to do. When the leadership of that agency is in flux, the evaluation process becomes unpredictable. Developers do not know which standards will apply, which tests will be required, or whether the person who approves their model today will still be in charge when they need approval for the next version. This uncertainty does not make AI development safer. It makes planning impossible, pushing companies to either rush releases before rules are finalized or delay innovation while waiting for clarity. Neither outcome serves the public interest.
The Geopolitical Pressure Cooker
The leadership gap at CAISI comes at a moment when the global AI race is accelerating. Chinese AI models have gained significant attention in the U.S. market, and the pressure on Washington to move quickly is intense. The U.S. government recently granted the UAE license-free access to advanced AI chips, signaling a major shift in export policy that could reshape the competitive landscape. Every month of leadership instability at CAISI is a month when the U.S. loses ground in setting the technical standards that will define how AI is tested, evaluated, and deployed worldwide. The agency’s work is not just about domestic safety. It is about establishing the protocols that other countries will adopt, which means whoever leads CAISI is also shaping the global governance of AI. An empty chair in Washington does not leave a vacuum. It leaves an opportunity for other governments to fill the gap with their own standards, which may not prioritize the same values.
The Structural Asymmetry of Accountability
The deeper problem is structural. The people who benefit from rapid AI deployment—investors, executives, early adopters—have clear incentives to minimize regulatory friction. The people who pay for AI failures—workers displaced by automation, communities exposed to algorithmic harm, citizens whose privacy is eroded—have diffuse interests and limited power. CAISI was designed to be the institution that rebalances this asymmetry by providing independent, technically grounded evaluation. But an institution that cannot keep its leadership is an institution that cannot fulfill its mission. The revolving door at the top sends a signal to everyone involved: this agency is not a priority but a political football, not a permanent pillar of the regulatory landscape. And when the people who are supposed to hold powerful AI systems accountable cannot hold onto their own jobs, the asymmetry only grows worse.
What the Next Director Will Inherit
The permanent director of CAISI will walk into an office with a desk, a staff, and a stack of problems that no single person can solve alone. They will need to establish credibility with both the AI industry and the public, two constituencies that often want opposite things. They will need to navigate the politics of a government that is simultaneously pushing for AI dominance and worried about AI risks. They will need to build evaluation protocols that are rigorous enough to catch dangerous capabilities but flexible enough to keep up with rapid technical progress. And they will need to do all of this while knowing that their predecessor lasted 90 days and the one before that lasted less than a week. The job comes with enormous responsibility, limited authority, and a high probability of burnout. The question is not whether the next director will struggle but whether the system will allow them to succeed.
The Real Barrier Is Not Technical
The greatest barrier to effective AI testing is not a lack of technical tools or scientific knowledge but the absence of institutional stability. The researchers at CAISI know how to evaluate models. They have the methods, the data, and the expertise. What they lack is consistent leadership that can translate their technical work into policy, secure the agency’s budget, and defend its independence against political pressure. Every time a director leaves, the agency loses institutional memory, strategic continuity, and the relationships with industry partners that take months to build. The technical work continues under acting leadership, but the planning that turns that work into lasting standards gets pushed further into the future. The AI industry does not need another safety pledge or another white paper on responsible development. It needs a government agency that can stay the course long enough to make its evaluations matter.