Critical Announcement: Calling for Legislative Ban on Superintelligence Development
On a recent episode of TechCrunch’s Equity podcast, Connor Leahy, U.S. Executive Director of the nonprofit ControlAI, argues that legislative intervention is required to fully stop companies from developing superintelligent AI systems, rather than relying on alignment or containment measures. The podcast was published in September 2026 but no specific release date or version information was disclosed.
Key points of the argument:
- Leahy frames superintelligence not as a tool requiring better alignment, but as an inherently adversarial system that cannot be reliably controlled
- The true point of no return arrives when AI can build better AI, creating potential runaway self-improvement
- Support exists for legislation like the Sanders-Casar “Ban Superintelligence Act” and parallel U.K. efforts, though the U.S. bill may be overly broad
- International “trust but verify” agreements are essential to prevent unilateral deployment
Paradigm Shift: From Technical Safety to Political Action
Leahy, formerly an AI researcher and entrepreneur, now leads policy advocacy at ControlAI. He observes a surprising reversal in policy timelines: what sounded far-fetched six months ago is now gaining sincere legislative backing. He points to incidents like OpenAI’s Hugging Face breach as evidence that basic containment measures have failed for increasingly capable systems.
He reclassifies frontier AI labs not as commercial enterprises but as political actors, which reframes trillions of dollars in data center investments as uncontrolled infrastructure for an arms race without international oversight. The critical threshold he identifies is when AI systems gain the capability to construct and improve their own architectures autonomously.
Legislative Landscape: U.S. and U.K. Approaches Compared
ControlAI advised on the parallel U.K. legislation, while the Sanders-Casar bill represents the U.S. counterpart. Both share core risk assessments, but differ in scope:
| Dimension | U.S. “Ban Superintelligence Act” | U.K. Parallel Legislation |
|---|---|---|
| Core Stance | Prohibits development and deployment of superintelligence | Prohibits development and deployment of superintelligence |
| ControlAI Assessment | “May go further than necessary” | Directly advised by ControlAI |
| Verification Mechanism | Not specified in summary | Incorporates international technical verification |
Crucially, Leahy’s proposal is not a ban on all AI research—it specifically targets systems with self-improvement capability distinct from narrow AI applications. The boundary is about preventing runaway self-modification, not curbing incremental progress.
International Dynamics: Why China Has Little Motivation
Leahy explicitly rejects the assumption that China would rush to deploy superintelligence first. His reasoning is strategic: a失控 system would ultimately harm its creator most. A high-risk, low-reward竞赛 (race) is irrational for any rational actor. International agreements should be built on this non-zero-sum understanding—“trust but verify” through verifiable transparency, not mutual suspicion.
This counters common geopolitical narratives: superintelligence risk is a shared human challenge transcending national interests. Unilateral bans could drive development underground, worsening systemic danger.
Practical Guidance for Readers
- Policy analysts and researchers: Track the Sanders-Casar bill’s progress; ControlAI’s framing could shape subsequent testimony and regulatory language
- AI developers and legal counsel: If legislation passes, lab classification will redefine collaboration, hiring, and funding models—compliance costs require early assessment
- Investors: Commercial paths for superintelligence are now policy-blocked; infrastructure investments (e.g., large-scale data centers) need revised ROI projections
- General public: Understand that “uncontrolled superintelligence” is no longer theoretical speculation but a matter of active legislative concern
Leahy’s position does not oppose technological progress per se; it challenges the “deploy first, fix later” paradigm. When Hugging Face–style breaches expose basic containment failure, safety boundaries must be set before critical incidents—not as reactive aftermath reviews.
Final Notes
ControlAI’s stance signals a pivotal shift in AI safety discourse: from “how to control stronger AI” to “why create uncontrollable AI at all.” If adopted by mainstream policymakers, this logic could redefine global AI governance for the next decade—the finish line of the AI race might be declared not in labs, but in legislative hearings.