Google DeepMind says AI now writes more than 75% of its production code, but its own safety report found no statistically significant acceleration in the pace of AI research and development. The figures show substantial use of AI inside the lab, not a self-running cycle of model improvement.
Google DeepMind Watch analysis
What happened
The Next Web reports that Google DeepMind’s Gemini 4 Argon safety and security report surveyed 12 senior leaders and researchers. It says the share of fully delegated pull requests rose from 7.5% in April to 11.9% in September. About 88% still needed repeated human steering or included code written by a person, and code changes remained subject to human review.
The report also compares the rate at which Gemini improves on an internal index of 92 benchmarks. The monthly rise went from 0.091 points before 1 April to 0.118 afterwards, a difference the report calls not statistically significant. Argon scores above that trend, but Google suspects increased investment and training compute, rather than AI-driven acceleration, explain the gap. The Next Web’s report also quotes Google co-founder Sergey Brin saying researchers and models now work together on the next version.
Why it matters
AI is already doing a large share of some coding work at Google DeepMind. That is a meaningful change to research workflows, but it is not the same as AI independently choosing research directions, designing experiments and improving future models at an accelerating rate. The report says people remain key to those tasks, while Argon solves only around half of the challenges in the lab’s internal test of AI research tasks.
Those distinctions matter as claims about recursive self-improvement gather pace. More AI-written code is evidence of adoption; it is not, by itself, evidence that AI is speeding up the whole research process.
Our read
The striking number is the code share. The more consequential claim, that AI is already accelerating AI development, gets a cooler reception from Google DeepMind’s own report. That is not a contradiction so much as a reminder that writing code is one part of research, not the whole scientific enterprise. For now, the evidence points to powerful assistance with people still steering the work.
What to watch
- Whether Google DeepMind publishes later benchmark comparisons showing a statistically significant change in improvement rates.
- Whether the share of fully delegated pull requests continues to rise, and what human review still catches.
- Whether the company reports measurable gains in research speed, rather than code production alone.
Discussion spark: What evidence would convince you that AI is accelerating AI research itself, rather than simply writing more of the code?
Sources and evidence
- Sergey Brin says Google’s AI now helps build its next models – The Next Web (17 October 2026, 19:16 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind