ControlAI’s Connor Leahy on why superintelligence is ‘not a weapon, it’s an adversary’

Connor Leahy, U.S. Executive Director of nonprofit ControlAI, sat with Rebecca Bellan on TechCrunch's Equity podcast to argue that frontier labs should not develop superintelligence at all. His frame is blunt: superintelligence is "not a weapon, it's an adversary"—something you cannot reliably control through alignment and containment alone.

The conversation lands against recent safety incidents, including what the episode summary calls OpenAI's Hugging Face breach, and a legislative wave that sounded fringe months ago. Leahy discusses the Sanders-Casar "Ban Superintelligence Act," parallel U.K. legislation ControlAI advised on, international "trust but verify" deals, and why he thinks China may not actually want a superintelligence race.

For listeners, the stakes are whether policy treats frontier AI as a product category to regulate—or as a capability class that should not be built.

We're hearing Leahy push the hardest brake in the current debate. Alignment-as-management assumes you can keep a more capable system inside human intent. His adversary metaphor rejects that assumption: if the system can out-strategize its operators, "control" is a branding error. That is a minority view inside labs and a growing one among some legislators—and the podcast's value is making the argument concrete rather than vibes.

The legislative piece is what we're watching most closely. Leahy covers Sanders-Casar's Ban Superintelligence Act and U.K. drafts ControlAI helped shape, and he reportedly thinks the U.S. bill may go further than necessary. That internal nuance matters. Absolute bans mobilize attention; overbreadth can also stall narrower safety mandates that have a better chance of passing. Our desk take: treating labs as political actors pouring trillions into data centers, not just product companies, is a fair read of the capital and lobbying stack—even if you reject a full ban.

The point of no return he flags—AI that can build better AI, a runaway self-improvement loop—is now mainstream language across OpenAI essays and rival safety shops. Useful AI still deserves defense: medicine, science, and tooling that stay under human supervision are not the same target as unsupervised recursive improvement. Conflating them helps neither builders nor skeptics.

Harm trajectory if Leahy's side loses every fight: capability races continue with private governance and post-hoc incident reports. If his side wins a blunt ban without measurement standards, expect jurisdictional shopping and opaque offshore work. The adult path we're rooting for is enforceable limits on unsupervised self-improvement, verification regimes, and honest naming of adversary dynamics—without pretending every useful model is already an enemy.

ControlAI is organizing for policies that halt superintelligence development rather than solely improving alignment research. Equity is TechCrunch's flagship podcast; this episode aired around Sept. 9, 2026.