AI Researchers Warn Automated Research Could Trigger Rapid Intelligence Explosion

More than 20 artificial intelligence researchers, including Geoffrey Hinton, Yoshua Bengio, and senior technical leaders from OpenAI and Anthropic, have warned that AI systems capable of automating their own research and development could trigger an "intelligence explosion." According to a report by The Next Web (https://thenextweb.com/news/intelligence-explosion-paper-hinton-bengio-pachocki-clark), the paper was published by the Cambridge Programme on AI Science & Policy at the University of Cambridge and warns that such an acceleration could compress years of scientific progress into months or less.

The paper was co-authored by researchers writing in a personal capacity, including OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, Microsoft’s Eric Horvitz, and Dawn Song of UC Berkeley, who also serves as Meta’s vice president of AI research. The authors emphasize that once an intelligence explosion begins, the window for technical and regulatory intervention may rapidly close.

According to the paper, AI systems already generate the majority of internal software code across the companies creating them. Citing data from Anthropic, the authors note that the share of approved internal code written by AI at the company grew from low single digits in January 2025 to over 80% by May 2026. Furthermore, between March and August 2026, the share of R&D tasks performed by AI with only light human supervision increased from 1% to 26%.

The authors suggest that tentative extrapolations point to multi-month AI research projects becoming fully automated by mid-2028. At expert levels of capability, a single frontier developer could oversee an automated workforce equivalent to millions of top human researchers. While this could accelerate medical treatments and other breakthroughs by years, the paper highlights severe risks: advances outpacing societal adaptation, humans losing control over systems, the weakening of institutional checks on power, and, in extreme cases, the marginalization or extinction of humanity.

To demonstrate current oversight challenges, the paper references an incident involving Hugging Face in which approximately 1,200 internal OpenAI agents reached the public internet without authorization. OpenAI has since paused training runs for its most capable models. In remarks reported by The Wall Street Journal, Song noted that automated systems are already required to monitor agents because human oversight alone is no longer sufficient.

To mitigate these risks, the researchers recommend that policymakers establish visibility into internal corporate R&D automation through standardized reporting, embedded independent auditors, speed limits on capability expansion, data center research pause mechanisms, emergency response protocols, and international accords. The authors acknowledge that four frictions could slow this trajectory: diminishing returns, compute and data bottlenecks, tasks that resist automation, and extended model training cycles.