AI companies have been talking about superintelligent AI like it’s inevitable, but recent safety incidents like OpenAI’s Hugging Face breach are demonstrating the potential dangers of deploying AI systems that are more capable than humans. Connor Leahy, an AI researcher, entrepreneur, and now the U.S. Executive Director of nonprofit ControlAI, is pushing for a far more radical approach: stopping companies from developing superintelligence altogether.
On a recent episode of TechCrunch’s Equity podcast, Leahy joined host Rebecca Bellan to explain why he thinks the risks have become too great to manage through alignment and containment alone. What sounded far-fetched six months ago, he argued, is now being backed by a wave of new legislation.
The OpenAI Hugging Face breach
Leahy’s argument starts with recent safety incidents that demonstrate what can go wrong when AI systems outpace human oversight. He points to OpenAI’s Hugging Face breach as a specific example of the potential dangers of deploying AI systems that are more capable than humans.
The breach is part of a broader pattern. AI companies talk about superintelligence as though it is inevitable, but the systems themselves are becoming harder to reliably control. That is the gap ControlAI wants to close — not with better safety research, but with a hard stop.
The question at the center of the episode is blunt: what happens when we can’t reliably control what these systems do?
What ControlAI wants
ControlAI is not in the business of tweaking safety protocols. The nonprofit is pushing for companies to halt superintelligence development entirely. Leahy explains why he thinks alignment and containment — the two dominant approaches to AI safety — are no longer enough.
He told Bellan that the risk has become too great to manage through those methods alone. The shift in his thinking, and in the broader conversation, is notable. Six months ago, a proposal to ban superintelligence development would have sounded far-fetched. Now, he says, it is being backed by new legislation.
The episode does not detail how alignment or containment work in practice. What it does make clear is that Leahy considers both approaches insufficient for the scale of the threat.
The legislation wave
Leahy described a position that sounded far-fetched six months ago. Now, he says, it is gaining support from a wave of new legislation. Public incidents involving AI systems have made the abstract concrete. When a company like OpenAI — one of the most prominent AI labs in the world — suffers a security breach on a platform like Hugging Face, the question of control stops being academic.
ControlAI’s position is straightforward: the risks have become too great to manage through alignment and containment alone. The nonprofit is pushing for a stop to superintelligence development, not more research into making it safe.
A shift in the conversation
The most striking part of Leahy’s argument is how fast the ground has moved. Six months ago, his position would have been dismissed as alarmist. Now it is the subject of legislative proposals.
That does not mean the debate is settled. AI companies continue to push forward. But the terms of the discussion have changed. The question is no longer just about making superintelligence safe. It is also about whether to make it at all.
Leahy’s answer is clear. He is not arguing for more research, more oversight or more testing. He is arguing for a stop.
Whether the industry listens is another matter. But the fact that the question is being asked is itself a shift.
ControlAI’s position may not win the day. But it has changed what counts as a serious position to hold.
Key facts box
- Guest: Connor Leahy, U.S. Executive Director of ControlAI
- Host: Rebecca Bellan, TechCrunch’s Equity podcast
- Incident cited: OpenAI’s Hugging Face breach
- ControlAI’s position: Stop superintelligence development entirely
- Timeline: Position sounded far-fetched six months ago; now backed by new legislation
Comparison: alignment, containment, and prohibition
| Approach | What it tries to do (general background) | ControlAI’s view |
|---|---|---|
| Alignment | Make AI systems want what humans want | No longer sufficient |
| Containment | Keep AI systems in a controlled environment | No longer sufficient |
| Prohibition | Stop superintelligence development altogether | ControlAI’s proposed approach |
The episode is available now on all major podcast platforms. Subscribers can follow Equity on X and Threads at @EquityPod.
Source: techcrunch.com
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.