Geoffrey Hinton and top scientists at OpenAI, Anthropic and Microsoft call for urgent oversight of AI that improves itself

The paper. A white paper released Monday, Sept. 28, warns that AI that automates its own development could set off an "intelligence explosion": each new system builds a better one, faster than people can check the work. In one scenario the paper describes, a year of AI progress happens in weeks. The names. The authors include Geoffrey Hinton, OpenAI's chief scientist, Microsoft's chief scientific officer and Anthropic co-founder Jack Clark. The evidence. They point to evidence that systems at OpenAI and Anthropic already do some of those companies' own research. They warn the "window for action may close" once AI starts improving itself. The limits. They also list four limits that could stop it: computing capacity, automation challenges, diminishing returns and slow training runs.

What the labs' own numbers say AI already does:

• Anthropic's merged code written by AI: low single digits in January 2025, over 80% by May 2026.
• Anthropic research led by AI under light human supervision: 1% in March 2026, 26% by August 2026.
• OpenAI's research arm ran 3.1 AI agent-workdays per human workday by mid-August 2026. More than half of its successful 4-to-8-hour agent tasks still needed a human to step in.