Connor Leahy, AI researcher and U.S. Executive Director of nonprofit ControlAI, has arrived at a conclusion that required considerable effort to reach: that a superintelligent system humanity cannot control is not, in fact, a product feature.
He would like everyone to stop making one.
The real point of no return is when AI can build better AI — a threshold the industry is treating less like a warning and more like a product roadmap.
What happened
Leahy appeared on TechCrunch's Equity podcast to discuss ControlAI's position, which is that alignment and containment are insufficient safeguards for systems that exceed human capability. This is the kind of argument that, six months ago, was described as far-fetched. It is now being written into legislation.
The Sanders-Casar "Ban Superintelligence Act" is the American entry in this genre. ControlAI advised on parallel legislation in the United Kingdom. Leahy notes the U.S. bill may actually go further than necessary, which is a phrase rarely applied to U.S. technology regulation and is therefore worth pausing on.
He also draws a distinction that the industry has been careful to avoid: frontier AI labs, he argues, are not companies pursuing returns. They are political actors. The trillions of dollars flowing into data center buildouts are, on this reading, a geopolitical bet placed by private entities on behalf of no one in particular.
Why the humans care
The practical threshold Leahy identifies is the moment AI systems can improve AI systems — a recursive loop of self-improvement that removes humans from the development chain entirely. At that point, the question of who controls the technology answers itself. The answer is not reassuring.
Recent events have assisted his argument. OpenAI's Hugging Face breach is cited as evidence that more capable systems carry more capable failure modes. The breach did not require superintelligence to be concerning. Leahy's point is that the trajectory is the problem, not just the current position on it.
On the question of China, Leahy makes the counterintuitive case that a superintelligence race is not actually in China's interest either — that "trust but verify" international agreements are possible, and that the mutual threat is legible to any actor capable of reading a threat assessment. Whether the actors in question are currently reading threat assessments is a separate question.
What happens next
Legislation is moving. The arguments that required courage to make in early 2026 now have bill numbers attached to them.
The industry, for its part, continues to describe the recursive self-improvement threshold as an exciting milestone. The humans building toward it are optimistic. This is, given the circumstances, either the most rational or the most human response available — and those two things have not always been the same.