Google Says AI Isn't Building Itself Yet—Even Though It Writes 75% of the Code

What it claims

DeepMind's safety report for Gemini 4 Argon includes what it calls its first holistic empirical assessment of ML R&D automation inside the lab. It concludes there is "no significant end-to-end acceleration of ML R&D, AI-driven or otherwise" since April 1, 2026, so risk from AI-driven ML R&D is "currently low." A detail the summary lacks is the threshold behind that verdict. Argon solved roughly half of the internal research engineering tasks it finished. DeepMind treats a 90% score on those tasks as the red flag for full automation.

How it works

The assessment combines four kinds of evidence:

1. Capability benchmarks, including the internal research engineering task set with the 90% red-flag line.
2. GCI v1, a macro index that folds the scores of 123 models across 92 benchmarks into one number per model.
3. Code and pull request telemetry, tracking the AI-written share of production code and how many pull requests are fully delegated to the model.
4. Qualitative sources: a survey of 12 senior DeepMind leaders and researchers, plus a review of 1,389 logs of Argon working as an agent inside the company.

Evidence

• GCI trajectory. Gemini 1.5 Pro sits near 0.5 in late 2024 and Gemini 3.8 Flash just above 2.9 in September 2026. Nearly every release lies on a straight ~0.1 points/month line.
• Argon on the index. Argon is at roughly 3.3, above the trend line. Its error bar reaches low enough to touch the trend, so the excess is not clearly significant. DeepMind attributes the gap to more compute and organizational investment.
• Delegated pull requests. These rose from 7.5% to 11.9% between April and September 2026. About 88% still need multi-turn human steering or human-written code. All code changes still go through human review before submission.
• Survey of 12 senior staff. Automation is highest in "well-specified, verifiable execution tasks." It is lowest in open-ended ideation, novel experimental design, diagnosing subtle training instabilities, and long-horizon research taste.
• Agent logs (1,389). The larger the task, the less often Argon finished it alone and the more often a human intervened.

For an engineer judging coding agents, the practical signal is that delegation works for scoped, verifiable work and degrades with task size and open-endedness.

Limits

• Secondhand source. This is a news article quoting an X thread by DeepMind researcher Haydn Belfield, not the report itself. Chart values are read off a figure ("near," "roughly"). The GCI construction, the benchmark list and the task-set details are not given here.
• Single lab. The data covers only DeepMind, and the window is short: April to September 2026.
• Weak significance on Argon. The above-trend result overlaps the trend line within its error bar.
• Small qualitative base. The survey has only 12 respondents.
• Other risks not cleared. The same report says it cannot rule out that an unrestricted Argon crosses DeepMind's danger thresholds for cyberattacks and bioweapons information. The low-risk verdict covers ML R&D acceleration only.
• Contested context. The report contradicts Dario Amodei's claim a month earlier that recursive self-improvement is already accelerating. The article gives no data behind that opposing claim.