DeepMind Safety Report Finds No AI-Driven R&D Acceleration

Google DeepMind said in its Gemini 4 Argon safety report that it found no evidence that internal AI research and development speed had significantly accelerated, according to the Korean tech outlet digitaltoday.co.kr. The company also said that more than 75% of production code inside DeepMind is currently written by AI.

DeepMind said that since April 1, 2026, there was no statistically significant figure in the speed of AI performance progress, and that its own capability index rose by 0.1 points per month. The company used an internal metric called GCI v1, which combines 123 AI models and 92 benchmarks.

On GCI v1, Gemini 1.5 Pro was around 0.5 points in late 2024, and Gemini 3.8 Flash exceeded 2.9 points in September 2026. Gemini 4 Argon was about 3.3 points, above the trend line, but DeepMind attributed this to expanded compute resources and increased organizational investment rather than AI-driven acceleration of development.

At DeepMind, the share of pull requests rose from 7.5% in April to 11.9% in September. About 88% of those pull requests still required separate instructions and direct human code writing, and the model could solve only about half of internal research engineering tasks on its own — far from the 90% level DeepMind regards as a warning sign of automation.

In a survey of 12 DeepMind senior executives and researchers, automation levels were highest for clearly defined and verifiable execution tasks and lowest for open-ended scientific idea generation, new experiment design, learning instability diagnosis and long-term research direction setting. The same report stated it could not rule out the possibility that an unrestricted Gemini 4 Argon could cross risk thresholds in the areas of cyberattacks or biological weapons-related information.