Google DeepMind looked for evidence that AI is speeding up the development of AI itself. It didn’t find it.
In the safety report for Gemini 4 Argon, its most powerful model, the lab said AI research is not moving faster even though AI now writes most of the code shipped by DeepMind engineers.
That runs against a scenario that has alarmed some AI executives and fueled calls for a slowdown. The concern is recursive self-improvement: one AI builds a smarter version of itself, which then builds an even smarter one, until progress moves beyond anyone’s ability to test, regulate, or switch it off. The report described that as “a situation where existing guardrails are severely inadequate.”
DeepMind says AI-driven ML R&D risk is currently low
Haydn Belfield, a Google DeepMind researcher involved in the assessment, wrote on X: “There is currently no significant end-to-end acceleration of ML R&D, AI-driven or otherwise. We therefore conclude that the risk from AI-driven ML R&D is currently low.”
ML R&D refers to machine learning research and development, the work of building new AI models.
In a thread posted on October 14, 2026, Belfield said the Gemini 4 Argon Safety & Security Report presented DeepMind’s first holistic empirical assessment of ML R&D automation and acceleration inside Google DeepMind, combining capability benchmarks, a macro capability index, data on AI-written code, and trajectory outcomes.
DeepMind tracks the issue because, as Belfield wrote, “AI could speed up frontier capability gains faster than safety, governance, and societal adaptation can keep pace.” More automation also means broader access for AI agents inside company systems, creating more room for a misaligned model, one that pursues goals its makers never intended, to cause trouble.
Internal capability index shows a steady pace of improvement
To measure progress, the team built GCI v1, an internal index that blends the scores of 123 AI models across 92 benchmarks into a single number for each model. Belfield compared it to an IQ score for AI.
“Frontier Gemini capabilities have improved at a remarkably steady pace of ~0.1 GCI points/month,” he wrote.
In the report’s chart, Gemini 1.5 Pro sits near 0.5 points in late 2024, while Gemini 3.8 Flash is just above 2.9 in September 2026. Nearly every release tracks a straight-line trend. Argon is the one point that rises above it, at roughly 3.3, although its error bar stretches low enough to touch the trend line.
Belfield wrote: “While Gemini 4 Argon sits above [the] trend, we trace this to increased compute and organizational investment rather than reflecting AI-driven acceleration.” Google unveiled Gemini 4 Argon on September 30.
AI writes most production code, but humans still control key steps
Most production code at DeepMind is now more than 75% AI-written. Over the same period, the share of fully delegated pull requests, bundles of code changes submitted for review before they go live, climbed from 7.5% to 11.9% between April and September.
Even so, Belfield said “~88% of pull requests still require multi-turn human steering or human code authorship—and code changes remain subject to human review before submission.”
Automation is strongest in execution, weakest in open-ended research judgment
A survey of 12 senior DeepMind leaders and researchers found that automation is highest in “well-specified, verifiable execution tasks” and lowest in open-ended scientific ideation, novel experimental design, diagnosing subtle training instabilities, and long-horizon research taste.
Put plainly, AI is good at carrying out instructions and weak at deciding which ideas are worth pursuing.
DeepMind also reviewed 1,389 logs of Argon operating as an agent inside the company. The larger the task, the less likely the model was to finish it on its own, and the more often a human stepped in.
Findings contrast with recent warnings from other AI leaders
The report arrived a month after Anthropic CEO Dario Amodei argued that recursive self-improvement is already accelerating. Sam Altman and Elon Musk backed that warning, and the reaction sent chip stocks down almost 6%.
Days earlier, former Anthropic engineer Jacob Coxon had left the industry, warning that AI could “kill us all by the end of the decade.”
DeepMind’s numbers cover only its own lab. The same report also said it cannot rule out the possibility that an unrestricted Argon could cross its danger thresholds for cyberattacks and bioweapons information.
