Google co-founder Sergey Brin stated that the company is entering a phase of recursive self-improvement, with Gemini models assisting in the development of their successors. Brin highlighted the collaboration between developers and the new Gemini 4 Argon model, which was recently released to the public.
Despite these claims, a Google DeepMind safety report suggests that significant end-to-end acceleration of AI research has not yet been observed. The report, based on interviews with 12 senior experts, classifies the risk of AI-driven research acceleration as currently low.
Data shows that AI generates over 75% of DeepMind's working code. While the share of fully automated pull requests increased from 7.5% in April to 11.9% in September 2026, approximately 88% of these changes still require human intervention for final approval.
Internal performance metrics for Gemini show a slight increase in improvement rates since April 1, 2026, rising from 0.091 to 0.118 points per month. However, Google attributes this growth to increased investment and computing power rather than autonomous AI acceleration.
Experts from Anthropic, OpenAI, and Google DeepMind estimate that full-scale recursive self-improvement could emerge within 3 to 12 months. Meanwhile, U.S. Representative Ro Khanna has proposed legislation to restrict self-improving AI technologies.
