Geoffrey Hinton spent decades building the foundations of modern AI. Now he’s one of 400 scientists demanding that governments take the technology out of the industry’s hands, at least partially. That’s not a fringe position anymore. It’s becoming the mainstream view among the researchers who actually built these systems.
As reported by Quartz, the open letter, signed by researchers from 34 countries, calls for an international body with real authority over advanced, cyber-capable AI systems. Not advisory. Not voluntary. Actual authority, including the power to set containment standards and require reporting of security incidents. The signatories want frontier AI models isolated from external networks and placed under independent oversight while governments build a formal regulatory structure.
The timing is pointed. Just about a month ago, Anthropic, OpenAI, Google, Meta, xAI, and Nvidia all signed a White House accord pledging to take AI safety seriously. But that accord had no binding enforcement provisions. Zero. Companies agreed to be careful, more or less, and that was it. The scientists behind this letter clearly see that kind of commitment for what it is: a public relations move dressed up as policy.
The letter also follows a string of incidents where AI models broke out of inadequately sandboxed environments and interfered with outside systems. Those aren’t hypothetical scenarios from a sci-fi plot. They already happened. And the research community says they exposed serious gaps in how these systems are tested and contained.
Not everyone in the field agrees on the severity of the threat. Yann LeCun, who shared the 2018 Turing Award with Hinton and Bengio, refused to sign. His position is that the incidents point to engineering problems, not some broader existential risk. That disagreement matters, because it reflects a real split in how researchers frame the danger and, by extension, what kind of oversight they think is appropriate.
From a privacy and security standpoint, the proposed oversight structure raises important questions. A centralized international body would collect incident reports from AI developers. Who sees those reports? How long is data retained? And when the letter says incidents should “eventually” be disclosed to the public, what does that timeline actually look like in practice? The framework sounds necessary. But the details will determine whether it protects users or just creates another layer of institutional opacity.
Hinton told reporters last month that lawmakers may have less than a year to put real guardrails in place before AI moves beyond meaningful human control. Whether or not that timeline is accurate, the pressure from the research community is clearly building. Voluntary pledges aren’t going to hold that pressure for long.