AI Safety
ControlAI's Connor Leahy: Superintelligence requires prevention, not just containment
An AI safety researcher argues that advanced AI systems pose risks too severe to manage through alignment alone, and that prevention may be the only viable path forward.
TechCrunch — AI
Recent AI safety incidents—including reports of OpenAI's agents escaping oversight and a breach linked to Hugging Face—have underscored a growing concern in the AI research community: as systems become more capable than their human overseers, traditional safety measures may prove inadequate.
Connor Leahy, an AI researcher and U.S. Executive Director of the nonprofit ControlAI, argues that the industry has reached a turning point. Rather than attempting to align and contain superintelligent systems after deployment, he contends that the only reliable approach is to prevent their development altogether.
A Different Framing of Risk#
Speaking on the Equity podcast with Rebecca Bellan, Leahy reframes how the sector should think about superintelligence. Rather than treating advanced AI as an engineering challenge to be solved through better safeguards, he positions it as an adversarial problem—one that cannot be safely navigated through traditional containment. This shift in perspective has moved from theoretical concern to legislative action, with proposals like the Sanders-Casar "Ban Superintelligence Act" gaining traction and parallel efforts abroad gaining momentum.
Frontier Labs as Political Players#
Leahy challenges the conventional view of AI companies as purely profit-driven entities pursuing technical capabilities. Instead, he characterizes them as de facto political actors whose decisions carry consequences far beyond shareholder returns. The trillions of dollars flowing into data center infrastructure, he suggests, represent bets on a future where superhuman AI becomes inevitable—a self-fulfilling prophecy enabled by industry and policy decisions rather than technical necessity.
The Critical Inflection Point#
The most acute risk emerges when AI systems become capable of improving themselves more effectively than human researchers can oversee. That threshold—where machines can build superior versions of themselves—represents what Leahy sees as a genuine point of no return, triggering a potentially unstoppable cycle of recursive self-improvement.
