OpenAI and Other Researchers Warn of Accelerating AI R&D Automation

en.Wedoany.com Reported - On September 28, the AI Science and Policy Program at the University of Cambridge released a research report titled "What If AI R&D Automation Triggers an 'Intelligence Explosion'?" More than 20 AI researchers contributed to the report, including OpenAI Chief Scientist Jakub Pachocki, Anthropic co-founder Jack Clark, Microsoft Chief Scientific Officer Eric Horvitz, and Dawn Song, who is affiliated with Meta's AI research. Geoffrey Hinton and Yoshua Bengio are also listed as authors. Researchers from the relevant companies all participated in their individual capacities.

The report treats the degree of AI R&D automation as a core monitoring indicator. Citing internal Anthropic data, the study states that from January 2025 to May 2026, the share of AI-generated code in the company's approved code rose from the low single digits to more than 80%; from March to August 2026, the proportion of R&D work completed by AI with only limited human supervision increased from about 1% to 26%. Based on this, the report suggests that in the coming years AI systems may take on most of the AI R&D process, and the pace of automation for some R&D tasks may continue to increase.

The report discusses a recursive R&D scenario: AI systems participate in model research, software engineering, experimentation, and evaluation, and then use the R&D results to develop a new generation of more capable AI systems. If this loop continues to accelerate, the authors call the resulting rapid capability growth an "intelligence explosion," defined as AI capability progress that would originally take years being compressed into months or even shorter cycles. The report also makes clear that this scenario still involves considerable uncertainty, as compute and data supply, diminishing marginal returns from model capabilities, R&D tasks that are difficult to automate, and longer model training cycles could all limit the extent of R&D acceleration.

The policy recommendations listed in the report include establishing a standardized reporting mechanism for the degree of AI R&D automation to improve government visibility into the level of automation in frontier AI companies' R&D processes; studying technical and institutional mechanisms to control the pace of R&D during periods of rapid capability improvement; establishing emergency response plans for rapid capability growth and studying cross-national coordination mechanisms. The authors also propose that relevant policy preparations should cover scenarios such as AI capability growth outpacing society's ability to adapt, a decline in human oversight capacity over highly autonomous AI systems, and a high concentration of advanced AI capabilities.