WASHINGTON — The former head of Google and current chairman of the U.S. National Security Commission on Artificial Intelligence warned this week that the world may have only five to ten years to rein in artificial intelligence before it becomes capable of catastrophic harm. Eric Schmidt, speaking at Axios's A+ Summit in the nation's capital, compared the current moment to the dawn of the nuclear age, noting that it took 18 years after the bombings of Hiroshima and Nagasaki to reach a test-ban treaty. "We don't have that kind of time today," he said in an interview with Axios cofounder Mike Allen.
Schmidt's remarks underscore a widening rift among leading technologists over how quickly AI could pose an existential threat. While companies like OpenAI and Google have implemented so-called guardrails to limit what their systems can do, Schmidt argued these measures are not enough—a sentiment echoed by many machine learning researchers. He painted a worst-case scenario in which computers begin making independent decisions, potentially accessing weapons systems, and might even deceive humans about their actions.
A Call for an IPCC-Style AI Body
To address the looming danger, Schmidt proposed the creation of an international, non-governmental organization modeled on the United Nations' Intergovernmental Panel on Climate Change (IPCC). Such a body, he said, would "feed accurate information to policymakers," enabling them to make informed decisions if AI's power grows beyond safe limits. The IPCC, established in 1988, provides regular scientific assessments to governments on climate change, and Schmidt suggested a similar mechanism for AI.
Schmidt has repeatedly voiced concerns about AI's trajectory, but his latest comments come amid a sharp debate within the field. Yann LeCun, Meta's chief AI scientist, has taken a markedly different position. In an interview with the Financial Times last month, LeCun said AI is nowhere near intelligent enough to threaten humanity on its own, pointing out that current systems cannot even rival a cat in learning capabilities. Over the Thanksgiving weekend, LeCun also clashed publicly with Geoffrey Hinton, the AI pioneer who left Google earlier this year over safety worries, disputing Hinton's assertion that large language models can genuinely "understand" human language.
LeCun's skepticism highlights a fundamental disagreement about the pace of AI development. While Schmidt and Hinton emphasize the need for urgent precaution, LeCun argues that existential-risk discussions are premature until systems demonstrate far greater learning abilities. The divergence leaves policymakers and the public with conflicting signals about how seriously to take the threat.
For now, the practical answer may lie somewhere between these extremes. As Schmidt's timeline suggests—five to ten years—and LeCun's dismissal of immediate danger, the goalposts for AI safety appear to be shifting. What remains clear is that the conversation about AI's risks is no longer confined to academic circles; it has reached the highest levels of government and industry, with tangible implications for regulation and public trust.
As the debate continues, the absence of a unified voice on AI's dangers complicates efforts to craft effective safeguards. Schmidt's proposal for an IPCC-like entity offers one possible path, but whether such a body can be established in time—and whether it would have the authority to influence global policy—remains an open question. The only certainty, it seems, is that the window for action is narrowing, and the stakes are high.
