ControlAI urges a halt to superintelligent AI development
ControlAI presses for a stop to superintelligence work
Connor Leahy, the U.S. Executive Director of nonprofit ControlAI, has argued that companies should stop developing superintelligent AI rather than rely solely on alignment and containment measures. He made the case during an episode of TechCrunch’s Equity podcast hosted by Rebecca Bellan.
The discussion centres on AI systems that could become more capable than humans and on the question of whether their actions can be controlled reliably. Leahy, an AI researcher and entrepreneur, said the risks have grown too great to manage through the safety approaches commonly proposed for advanced systems.
Safety incidents sharpen the debate
AI companies have often presented superintelligent AI as an expected destination for the industry. The podcast examines the implications of deploying systems whose capabilities exceed those of people, particularly when operators cannot reliably control what the systems do.
Recent safety incidents have made that concern more immediate. TechCrunch cited OpenAI’s Hugging Face breach as an example of the potential dangers surrounding the deployment of increasingly capable AI systems. The episode uses that context to ask whether technical safeguards alone are sufficient as capabilities advance.
From alignment to limits on development
Leahy’s position goes beyond calls for stronger alignment practices or better containment. ControlAI is advocating a more radical intervention: preventing companies from developing superintelligence in the first place.
He said an argument that might have seemed far-fetched six months earlier is now receiving support from a wave of new legislation. The article does not identify the measures, but frames their emergence as evidence that policy attention to the issue is increasing.
For businesses evaluating AI systems, the practical implication is to examine control, oversight and failure scenarios alongside promised capability gains, rather than assuming that more powerful models automatically create manageable operational value.