Google / Wikimedia Commons (Public domain)
An internal audit tied to the Gemini 4 Argon safety report finds steady gains, heavy human review, and no sign of AI speeding up its own development
Google DeepMind published its first comprehensive internal assessment of AI-driven machine learning research on October 14, 2026. The audit found no measurable acceleration in AI progress that can be credited to AI itself.
What the audit actually measured #
The assessment appears inside the Gemini 4 Argon Safety & Security Report. Researcher Haydn Belfield led the work.
The team tracked frontier Gemini capabilities with a metric it calls the General Capability Index, or GCI. Across 123 models and 92 benchmarks, that score has climbed by approximately 0.1 GCI points each month.
The report says thorough statistical analysis shows no significant structural speedup after April 2026. For public models, the monthly change before April 2026 was 0.091 points. After that date, it was 0.118 points. The report found no substantial difference between the two.
Gemini 4 Argon itself landed marginally above the expected trend line. DeepMind’s report suggests that bump mainly comes from more compute and organizational investment, not from the model accelerating its own development.
AI writes the code, humans still sign off #
AI tools contribute to writing over 75% of production code at DeepMind. Some 88% of code changes still need human review.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Fully delegated changes, where AI handles the work end to end, rose from 7.5% to 11.9%.
Surveys of 12 senior DeepMind researchers help explain the pattern. Automation works best on clearly defined tasks where the answer can be verified. It struggles badly with open-ended scientific ideation and complex planning.
The benchmark scorecard #
DeepMind also ran its agents through RE-Bench, an internal benchmark of 74 realistic machine learning engineering tasks. The Argon agent solved roughly half of them. Gemini 3.7 Flash managed 27%.
A separate review of 1,389 Argon agent runs found that success rates fell notably on complex tasks when no human was involved.
What this means for the AI industry #
The share of fully delegated code changes is one trend to watch, since its move from 7.5% to 11.9% is the closest thing to an automation trend line in the report. A sustained break above roughly 0.1 GCI points per month, without a matching jump in compute, would be the signal the self-improvement camp has been waiting for. For now, DeepMind’s own ledger shows a steady climb and a lot of humans still reading the code.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our