Connor Leahy, the AI researcher now serving as US Executive Director of the nonprofit ControlAI, is arguing that frontier labs should be legally barred from building superintelligent systems at all. Speaking on TechCrunch's Equity podcast, Leahy said superintelligence is "not a weapon, it's an adversary," and that alignment and containment are no longer sufficient tools to manage the risk. His organization is backing the Sanders-Casar Ban Superintelligence Act in the US and advised on parallel legislation in the UK. Six months ago, Leahy said, that position sounded far-fetched — now it has legislative sponsors.
The pitch is a hard stop, not a speed limit. ControlAI's position is that no lab, in any jurisdiction, should be permitted to train models beyond a defined capability threshold. Leahy told the podcast the US bill may go further than necessary but that the direction is correct. The UK version, which ControlAI shaped, targets the same category of frontier development with a lighter regulatory hand.
Leahy's framing of the industry is unusually blunt. He described frontier AI labs less as commercial enterprises chasing returns and more as political actors, arguing that the trillions of dollars being poured into data center buildouts represent a form of geopolitical positioning rather than ordinary capex. That framing matters because it changes who a regulator is negotiating with. A company optimizes for revenue; a political actor optimizes for capability and leverage.
Key facts
- 01Connor Leahy now serves as US Executive Director of ControlAI, a nonprofit lobbying to halt superintelligence development entirely.
- 02The Sanders-Casar Ban Superintelligence Act is moving through the US Congress, with parallel UK legislation that ControlAI advised on.
- 03Leahy frames frontier labs as political actors steering trillions of dollars into data center buildouts, not just companies chasing returns.
- 04He argues the point of no return is when AI can build better AI, triggering runaway self-improvement.
- 05The push follows recent incidents including OpenAI's Hugging Face breach, cited as evidence that containment is failing.
The technical inflection point, in Leahy's telling, is recursive self-improvement. He argues the real point of no return is when AI can build better AI, creating a potentially runaway cycle where each generation of models designs its successor faster than humans can supervise. Once that loop closes, external oversight becomes structurally impossible — which is why ControlAI wants the intervention to come before, not after, capability thresholds are crossed.
The immediate evidence Leahy points to is a string of safety incidents at frontier labs, including OpenAI's recent Hugging Face breach, which exposed the fragility of current containment practices. If deployed models with today's capability profile can be compromised or misused, the argument runs, the case for shipping something an order of magnitude more capable weakens sharply. AI Chat Daily covered a related concern last week, when an Anthropic safety lead put AI extinction odds above 10% this decade — a figure that would have been fringe two years ago.
On the international dimension, Leahy pushed back on the standard framing that a US halt would simply hand the frontier to China. He made the case for "trust but verify" agreements between major powers and argued it isn't actually in China's interest to build superintelligence either, given the same loss-of-control dynamics apply regardless of which government hosts the training run. That position, if it holds, cuts against the dominant industry lobby argument that any US restriction is a unilateral surrender.
The bill itself, sponsored by Bernie Sanders and Greg Casar, is the first US federal legislation to attempt a categorical prohibition rather than a disclosure or licensing regime. Prior AI bills at the federal level have focused on transparency, evaluations, or sector-specific deployment rules. A ban is a different instrument, and it will face a different fight — one where the target isn't a specific harm but a specific class of system.
The counterargument from the labs is familiar. OpenAI, Anthropic, Google DeepMind, and xAI have all argued that safety research requires continued access to frontier models, that a halt would push development into less accountable jurisdictions, and that the risks Leahy describes remain speculative. None of the major labs has publicly endorsed anything close to the Sanders-Casar approach, and the trillion-dollar capex commitments already flowing into data center construction assume no such ceiling exists.
The ControlAI push lands at a moment when the regulatory conversation is fragmenting rather than consolidating. The EU AI Act is in enforcement, US state bills are proliferating, the UK is running a lighter-touch framework, and China has its own model-registration regime. A federal ban in the US would reshape that landscape overnight, forcing every major lab to redesign roadmaps and every hyperscaler to reassess data center pipelines that were sized for continued scaling. Whether the Sanders-Casar bill has the votes is a separate question from whether the political ground has shifted enough to make it thinkable — and on that second question, Leahy is clearly winning ground he didn't hold a year ago.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
