The Gemini 4 Argon safety report shows that AI writes more than 75% of the code at DeepMind but has not accelerated the pace of AI research.

Google DeepMind looked for evidence that artificial intelligence (AI) is accelerating the development of the technology itself, but found none. In the safety report for Gemini 4 Argon, the company’s most powerful model, unveiled on September 30, the first internal assessment of AI-assisted research concluded that the pace of development had not measurably increased, even though AI had taken on most of the code used in production.

This finding runs counter to the recursive self-improvement scenario that has worried many industry leaders: an AI system continually creating smarter versions of itself at a pace beyond humans’ ability to test, control, or shut it down. Haydn Belfield, a DeepMind researcher involved in the evaluation, said on X that there had been no significant increase in the pace of machine learning research and development (ML R&D) across the entire process, so the risks from this activity are currently low.

To measure progress, the research team developed GCI v1, an internal index that combines the results of 123 models across 92 benchmarks into a single score, similar to how human IQ is measured. According to Belfield, the capabilities of advanced Gemini models have improved steadily by about 0.1 GCI points per month. Gemini 1.5 Pro scored nearly 0.5 points at the end of 2024, while Gemini 3.8 Flash exceeded 2.9 points in September 2026, with most versions closely following the linear trend line. Argon was the only exception, at around 3.3 points, but its error bar still touched the trend line. DeepMind explained that this outperformance stemmed from increased computing resources and organizational investment, rather than acceleration driven by AI.

AI is good at execution, but not yet at choosing a direction

More than 75% of the production code at DeepMind is now written by AI. The share of code-merge requests delegated entirely to AI rose from 7.5% to 11.9% between April and September. However, about 88% of requests still require humans to guide the process over multiple rounds or write code directly, and every change must be reviewed by a person before it is submitted.

A survey of 12 senior DeepMind leaders and researchers found that automation is highest for execution tasks with clear requirements and verifiable outcomes. By contrast, AI remains limited in generating open-ended scientific ideas, designing new experiments, diagnosing subtle instability in model training, and choosing long-term research directions. A review of 1,389 records of Argon’s activity as an internal AI agent also found that the more complex the task, the lower its likelihood of being completed autonomously and the greater the need for human intervention.

DeepMind is monitoring this issue because AI could advance model capabilities faster than safety measures and governance mechanisms can adapt, while AI agents are increasingly being granted broad access to enterprise systems. The report was published a month after Dario Amodei, CEO of Anthropic, said that recursive self-improvement had begun to accelerate. Sam Altman and Elon Musk echoed the warning, while semiconductor stocks fell nearly 6%.

Even so, DeepMind’s data reflects only activity within the company’s research lab. The report also acknowledges that it cannot rule out the possibility that an unrestricted version of Argon could exceed the company’s identified risk thresholds for cyberattack capabilities and providing information related to biological weapons.