Google DeepMind's Belfield reports no measurable AI-driven acceleration of AI research; AI writes over 75% of production code but humans still review changes

Google DeepMind's Haydn Belfield published internal findings, via X and a safety report, saying AI now writes over 75% of production code but about 88% of code changes still need human oversight, and all are human-reviewed before submission. Fully AI-delegated changes rose from 7.5% to 11.9%. A new General Capability Index (123 models, 92 benchmarks) shows Gemini improving by about 0.1 points a month, with no statistically significant acceleration. A survey of 12 senior researchers is said to agree. The data is Google's own and the primary documents are not linked in the coverage.

This is a rare, quantified datapoint from inside a frontier lab on whether AI is speeding up AI research. It gives policymakers and safety teams something checkable to weigh against warnings from Hinton, Bengio and Pachocki. It also suggests that, for now, human review remains a real bottleneck in software work, which matters for anyone planning developer headcount or AI coding workflows. The practical impact is limited: it changes no products or deployments, and it describes one lab at one point in time.

The headline "self-improving AI isn't anywhere close" goes further than the evidence. A flat trend in one lab's own index over a short period does not rule out future acceleration, and Google rating its own risk as low is not independent verification. The 12-person survey is tiny, and "75% of code written by AI" says nothing about how hard or important that code is.

1. 01Publication of the full safety report and GCI methodology, so outsiders can check the 0.1 points a month figure and the 1,389-run Argon data
2. 02Whether the fully AI-delegated share of code changes keeps rising beyond 11.9% in DeepMind's next update, and whether other labs publish comparable figures
3. 03Independent or rival results on RE-Bench (Argon at about half of 74 tasks) and on GCI trends for newer models, showing whether growth stays near linear or bends upward