Google DeepMind has found no statistically significant evidence that artificial intelligence is speeding up the development of AI itself, despite the lab now using AI to write more than 75% of its production code.
The conclusion appears in the safety and security report for Gemini 4 Argon, Google’s most powerful model, and addresses concerns about recursive self-improvement. That scenario describes an AI system helping create a more capable successor, which then accelerates the process further. In the most extreme version, progress could outpace the ability of researchers and regulators to test systems, impose safeguards or shut them down.
According to Decrypt, DeepMind’s assessment found that the pace of machine-learning research and development has not accelerated since April 1, 2026. Haydn Belfield, a Google DeepMind researcher involved in the work, said the lab therefore considers the current risk from AI-driven machine-learning research to be low. The assessment applies to the lab’s measured experience and does not establish that such risks are impossible elsewhere or in the future.
A steady capability trend
To track changes in AI performance, DeepMind created GCI v1, an internal capability index. The index combines results from 123 AI models across 92 benchmarks into one score for each model, providing a common way to compare progress across releases.
The report says frontier Gemini capabilities have risen at a steady rate of about 0.1 GCI points per month. Its chart places Gemini 1.5 Pro at around 0.5 points in late 2024 and Gemini 3.8 Flash slightly above 2.9 in September 2026. Most releases followed a broadly straight-line trend.
Gemini 4 Argon, which Google unveiled on September 30, was positioned at roughly 3.3 points, above that trend. However, the uncertainty range around its result extended low enough to meet the established trajectory. DeepMind attributed the higher result to increased computing resources and organisational investment, rather than to AI systems accelerating the research process itself.
The distinction matters because faster AI capability gains could create a gap between technological progress and the development of safety controls, governance and broader social responses. More automation can also require AI agents to receive wider access to internal systems, potentially increasing the consequences of a model pursuing objectives its developers did not intend.
AI writes code, but humans still direct the work
DeepMind’s software data shows substantial adoption of AI tools without demonstrating full automation. More than 75% of production code at the lab is now written by AI. The proportion of fully delegated pull requests—groups of code changes submitted for review before release—increased from 7.5% in April to 11.9% in September.
Even so, about 88% of pull requests still required several rounds of human guidance or direct human code writing. Code changes also remained subject to human review before submission. The figures suggest that producing code is only one part of the research process; people continue to define tasks, assess results and decide whether proposed work should move forward.
A survey of 12 senior DeepMind leaders and researchers found that automation was strongest in tasks that were clearly specified and easy to verify. It was weakest in open-ended scientific idea generation, designing new experiments, identifying subtle training problems and applying long-term research judgment.
DeepMind also examined 1,389 records of Argon operating as an agent within the company. The analysis found that as assignments became larger, the model was less likely to complete them independently and human intervention became more frequent. Argon completed roughly half of the internal research engineering tasks it attempted to finish. That result remains well below the 90% level DeepMind treats as a warning sign for full automation.
The wider debate over recursive self-improvement
DeepMind’s findings arrive amid a disagreement over how quickly AI research may become self-reinforcing. One month earlier, Anthropic Chief Executive Dario Amodei argued that recursive self-improvement was already accelerating. Sam Altman and Elon Musk endorsed the warning, and chip stocks fell almost 6% at the time, according to the source material.
The report does not dismiss all potential risks. DeepMind said its analysis covered only its own laboratory and could not rule out the possibility that an unrestricted Argon might cross its danger thresholds relating to cyberattacks and information about bioweapons.
The central result is narrower: despite widespread AI involvement in software production, DeepMind found no significant end-to-end acceleration in machine-learning research and development after April 1. AI appears highly useful for carrying out well-defined work, while humans remain responsible for much of the direction and judgment behind longer-horizon research.
Moneyiar market picture
Moneyiar market data at 21:24 on October 15, 2026, showed DeepBook Protocol (DEEP) at $0.0230012331, down 4.42% over 24 hours but up 14.09% over seven days. AI Rig Complex (ARC) was $0.0387741802, up 9.42% over 24 hours and down 36.9% over seven days.
At 21:23, aelf (ELF) was priced at $0.0625597756, down 0.21% in 24 hours and 3.47% over seven days. Propy (PRO) stood at $0.3719933611, up 2.25% on the day and 9.71% over the week.
