Google DeepMind Says AI Is Not Yet Accelerating AI Research, Even Though AI Writes 75 Percent of Its Code

Google DeepMind went looking for evidence that AI is speeding up the development of AI itself. It did not find any. In the safety and security report published alongside Gemini 4 Argon, the lab laid out its first holistic empirical assessment of machine-learning R&D automation inside the company, and researcher Haydn Belfield summarized the conclusion bluntly 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."

The question matters because recursive self-improvement — an AI building a smarter successor, which builds a smarter one still, until progress outruns anyone's ability to test or switch it off — is the scenario that has spooked the industry's own leadership. Anthropic CEO Dario Amodei argued a month ago that the acceleration is already underway, a warning Sam Altman and Elon Musk endorsed and one that knocked chip stocks down almost 6 percent at the time. DeepMind tracks the issue because, as Belfield put it, AI could speed up frontier capability gains faster than safety, governance and societal adaptation can keep pace.

To measure it, the team built GCI v1, an internal General Capability Index that blends scores from 123 AI models across 92 benchmarks into a single number per model. The result is a nearly straight line: frontier Gemini capability has improved at a remarkably steady pace of roughly 0.1 GCI points per month, from about 0.5 for Gemini 1.5 Pro in late 2024 to just above 2.9 for Gemini 3.8 Flash in September 2026. Argon is the one dot above the trend at roughly 3.3, but the report attributes that to increased compute and organizational investment rather than AI-driven acceleration.

The code numbers tell a similar story of automation without autonomy. Most production code at DeepMind is now more than 75 percent AI-written, and the share of pull requests fully delegated to AI rose from 7.5 percent to 11.9 percent between April and September. But about 88 percent of pull requests still require multi-turn human steering or human code authorship, and every code change is reviewed by a person before submission. Three quarters of the code coming from a model is not the same as the model running the lab.

Agent-level data points the same way. DeepMind reviewed 1,389 logs of Argon working as an internal agent: the bigger the task, the less likely the model finished it alone and the more often a human stepped in. On RE-Bench, an internal set of 74 realistic ML engineering tasks, Argon solved roughly half — well below the 90 percent threshold DeepMind treats as a red flag for full automation. A survey of 12 senior leaders and researchers found automation 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. In plain terms, AI executes instructions well and chooses which ideas deserve chasing poorly.

The caveats are real. This is one lab measuring itself, with first-party data and no external replication, and the same report notes it cannot rule out that an unrestricted version of Argon would cross DeepMind's danger thresholds for cyberattacks or bioweapons-related information. The finding is also not a claim that nothing is changing — a 75 percent AI-written codebase is a profound operational shift — only that the compounding loop, where models materially shorten the cycle of building better models, is not visible in the data yet.

The useful number to watch going forward is the slope. If the GCI index's 0.1-points-a-month line starts bending upward without a matching jump in compute and headcount, that is the fingerprint of recursive self-improvement, and it would show up in a chart like this long before it shows up in a press release. For now, the most consequential AI lab in Europe has published its own evidence that the feedback loop everyone fears — and everyone markets — has not started. That is either reassuring, or a reminder that the people most confident about imminent superintelligence have not shown their data.