Google’s DeepMind lab disclosed that more than three-quarters of the code shipped by its engineers is now written by artificial-intelligence systems. The proportion of pull requests that are fully delegated to AI rose from 7.5% to 11.9% between early April and September 2026. Despite this high level of automation, roughly 88% of code modifications still involve multi-turn human guidance or direct authoring, and every change undergoes human review before deployment.
No measurable acceleration in machine-learning research
The lab’s latest safety and security report, centered on its most powerful model Gemini 4 Argon, found no statistically significant increase in the speed of AI research after April 1 2026. An internal metric called the Gemini Capability Index (GCI) – which aggregates scores from 123 models across 92 benchmarks – has risen at a steady rate of about 0.1 points each month. Gemini 4 Argon briefly exceeded the trend line, but researchers attribute this outlier to additional compute resources rather than a self-reinforcing feedback loop.
Human steering remains essential for complex tasks
DeepMind examined 1,389 logs of Argon operating as an autonomous agent within the company. The data show that the larger and more open-ended a task, the more likely a human intervenes. A survey of senior DeepMind staff confirmed that AI excels at well-specified, verifiable execution but struggles with scientific ideation, novel experiment design, diagnosing subtle training instabilities, and long-term research direction. Consequently, most high-impact research still depends on human insight.
Safety implications and industry context
The report arrives amid heightened industry debate about recursive self-improvement – the notion that an AI could iteratively build increasingly capable successors, potentially outpacing oversight mechanisms. DeepMind’s findings suggest that, at least within its own operations, such a runaway scenario is not yet observable. Researchers concluded that the risk from AI-driven machine-learning R&D is currently low, though they caution that broader access to company systems could increase exposure to misaligned models.
Why it matters
Understanding the balance between automation and human control is critical as AI tools become integral to software development and research. DeepMind’s data indicate that while AI can handle routine coding tasks at scale, the creative and strategic dimensions of cutting-edge AI work still require human expertise. The absence of a measurable acceleration in self-improvement tempers fears of an imminent intelligence explosion, yet the continued reliance on human oversight underscores the importance of robust safety frameworks as AI capabilities expand.
Why it matters
The findings provide a rare, data-driven perspective on how AI is reshaping internal engineering workflows and the pace of scientific progress. They suggest that, despite impressive automation levels, the industry remains far from a scenario where AI independently engineers ever-more powerful systems. Policymakers, investors, and technologists can use this evidence to calibrate regulatory approaches and resource allocation, ensuring that safety measures keep pace with incremental but steady advances in AI capability.
