Google says self-improving AI is nowhere close
Internal DeepMind measurements show steady gains and no acceleration, AI already writes 75 percent of company code while humans review 88 percent of changes
Images
The General Capability Index (GCI v1) shows steady growth in Gemini capabilities with no statistically significant acceleration. | Image: Haydn Belfield / Google Deepmind
Haydn Belfield / Google Deepmind
An engineer at Google DeepMind says self-improving AI remains far off, and the company's own measurements are his evidence. Haydn Belfield says the lab has seen no measurable acceleration of AI research driven by AI itself — even as AI now writes more than 75 percent of the production code at Google, The Decoder reports.
The fine print undercuts the headline figure. About 88 percent of code changes still need human oversight or review, and every change is reviewed by a person before it goes in. The share of changes fully delegated to AI rose from 7.5 to 11.9 percent — movement, but of the incremental kind. Across 1,389 internal runs of Google's Argon agent, success without human input dropped sharply as tasks grew more complex. Belfield says humans remain essential for creative research, novel experiment design, diagnosing subtle training instabilities and long-term planning, and a survey of twelve senior DeepMind researchers confirmed the pattern: automation runs highest on well-defined, verifiable tasks and lowest on open-ended scientific ideation.
The benchmarks agree. On RE-Bench, DeepMind's in-house suite of 74 realistic machine-learning engineering tasks, Argon solves about half — up from 27 percent for Gemini 3.7 Flash — a jump the team attributes to more compute and organisational investment rather than AI-driven acceleration. A General Capability Index covering 123 models and 92 benchmarks shows Gemini capabilities growing at roughly 0.1 points per month, with no statistically significant speed-up. The team rates the current risk from AI-driven research automation as low and is building measurement infrastructure to track signs of acceleration going forward.
The assessment lands amid louder forecasts. More than 20 researchers, including Geoffrey Hinton, Yoshua Bengio and OpenAI's Jakub Pachocki, recently warned in a paper of a possible "intelligence explosion", arguing that AI systems could automate all AI research and development within a few years, compressing years of progress into months. Google's reply is a dataset: measurements taken on its own agents, graded on its own benchmarks, by the company with the most compute to defend. A race that advances 0.1 points a month is a race that scale wins — and no participant has more scale than the one reporting no acceleration.
For now, the share of code changes Google fully hands to its own AI stands at 11.9 percent.