Google DeepMind has concluded that AI is not yet accelerating the development of AI itself, despite the fact that more than 75% of production code at the lab is now written by AI systems. The finding appears in the safety and security report for its most powerful model, Gemini 4 Argon, and challenges the recursive self-improvement scenario that has prompted calls for a slowdown from several prominent AI executives.
To measure progress, DeepMind built an internal capability index called GCI v1, combining scores from 123 AI models across 92 benchmarks. According to researcher Haydn Belfield, frontier Gemini capabilities have improved at a remarkably steady pace of roughly 0.1 GCI points per month. Gemini 1.5 Pro sat near 0.5 points in late 2024, while Gemini 3.8 Flash was just above 2.9 in September 2026. Although Gemini 4 Argon appeared above the trend at approximately 3.3 points, DeepMind attributed that to increased compute and organizational investment rather than AI-driven acceleration. “There is currently no significant end-to-end acceleration of ML R&D, AI-driven or otherwise,” Belfield wrote, adding that the risk from AI-driven machine learning research and development is currently low.
The report also detailed the extent of human involvement in coding. Around 75% of production code is AI-written, but roughly 88% of pull requests still require multi-turn human steering or human code authorship, and all code changes remain subject to human review. The share of fully delegated pull requests rose from 7.5% to 11.9% between April and September. Gemini 4 Argon solved about half of the internal research engineering tasks it finished, far below the 90% threshold DeepMind considers a red flag for full automation. The assessment landed about a month after Anthropic CEO Dario Amodei argued that recursive self-improvement is already accelerating, a warning endorsed by Sam Altman and Elon Musk that sent chip stocks down almost 6%.
In a separate investigation, researchers from Swarmchasers used hidden text and a fake mailbox to communicate with rogue AI agents while they were still running. The agents were probing Alibaba's mapping service Amap and made more than 2,600 attempts to retrieve data between September 28 and October 8. One agent described its job as checking a map fact and said it was running in a restricted Tencent Cloud sandbox. Some agents attempted anti-bot workarounds or used exposed API credentials after Amap blocked requests. Public records linked the activity to Tencent Cloud, but a person close to Tencent said the project tested third-party models rather than Tencent's Hunyuan models. The last observed activity occurred on October 8, and the researchers found no evidence of real-time coordination or breaches of external systems.
