More than 20 leading artificial intelligence researchers published an urgent white paper on September 28, 2026, warning that AI systems automating their own development could trigger an “intelligence explosion”. The paper warns such rapid scale-up would leave governments with little time to respond, with one scenario suggesting a year’s worth of advances could occur in weeks.
The warning’s authors include Turing Award winners Geoffrey Hinton and Yoshua Bengio, along with OpenAI’s chief scientist Jakub Pachocki, Microsoft’s chief scientific officer Eric Horvitz, and Anthropic co-founder Jack Clark. The paper, titled “What if automating AI R&D triggers an intelligence explosion,” describes recursive self-improvement, where AI systems identify their own bottlenecks and implement architectural changes autonomously.
The core evidence rests on Anthropic self-reports showing AI’s share of approved code rose from low single digits to over 80% between January 2025 and May 2026. Meanwhile, R&D work autonomously completed with only high-level human supervision jumped from 1% to 26% between March and August 2026. OpenAI has set a goal of a fully automated AI researcher by 2028. Extrapolations suggest months-long research projects could be automated by mid-2028.
Hinton recently told US senators they had roughly a year to act on AI, while Bengio has urged the UN Security Council to license frontier AI. Their participation carries weight, as both are widely credited with laying intellectual foundations for today’s AI boom, though both have long warned about negative consequences including widespread job displacement, cyberattacks, and biological weapons risks.
The research coalition from Anthropic, OpenAI, Meta, and Microsoft publicly asks governments to oversee their own employers, representing rare internal pressure for external regulation. The paper, published by Cambridge’s Programme on AI Science & Policy, warns that automated research could outpace human control and raises the risk of “marginalization or extinction of humanity.”
The researchers call for mandatory government oversight including incident reporting, independent evaluators, safety testing requirements, and mechanisms to pause AI work. They acknowledge uncertainty about whether recursive self-improvement will actually occur, noting computing constraints, automation challenges, and diminishing returns could prevent such scenarios. However, they emphasize response time would be minimal if acceleration begins.
The paper concludes: “Once an intelligence explosion begins, the window for action may close.”
