AI is taking on a growing share of software work at Google Deepmind, but that does not yet mean it can drive its own research forward. Google engineer Haydn Belfield says the organization has seen no measurable acceleration of AI research driven by AI itself.
More code is automated, with people still checking the work
AI now writes more than 75 percent of production code, according to Belfield. But about 88 percent of code changes still need human oversight or review, and every change is reviewed by a human before submission.
The portion of changes fully delegated to AI has risen from 7.5 to 11.9 percent. That increase signals a real expansion in automation, while the review figures show how much human involvement remains part of the process.
These measures describe different parts of software development. AI can produce code without taking full responsibility for deciding what work matters, checking whether a change is sound, or handling the broader direction of a research project.
Complex tasks expose the limits of independent work
Across 1,389 internal Argon agent runs, success without human input fell sharply as tasks became more complex. The pattern suggests that automation works best when a task is clearly defined and its result can be checked.
A survey of twelve senior Deepmind researchers found a similar split. Automation was highest for well-defined, verifiable tasks and lowest for open-ended scientific ideation.
Belfield said people remain essential for creative research, designing novel experiments, diagnosing subtle training instabilities, and long-term planning. Those responsibilities call for judgment across uncertain problems, rather than simply completing a bounded task with a clear answer.
Benchmark gains do not yet amount to full automation
Deepmind evaluates Argon on RE-Bench, an in-house benchmark with 74 realistic ML engineering tasks. Argon solves about half of them. That is a substantial improvement over Gemini 3.7 Flash, which solved 27 percent, but it remains short of the near-perfect performance that full automation would require.
A benchmark result can show progress on a set of tasks without establishing that an AI system can independently carry out the wider research process. The remaining gap matters because research includes choosing worthwhile questions, interpreting uncertain results, and planning what to try next.
Measured capability growth remains steady
To track AI progress, the team created a General Capability Index (GCI) spanning 123 models and 92 benchmarks. Gemini capabilities are increasing by about 0.1 GCI points per month, with no significant change in the rate of growth.
The team attributes Gemini 4 Argon’s position slightly above the trend line to more compute and organizational investment, rather than to AI accelerating AI research. On that evidence, the team currently rates the risk from AI-driven research automation as low, while building measurement infrastructure to watch for signs of acceleration.
Researchers disagree about how quickly the shift could come
Belfield’s findings align with recent research, but there is no consensus across the field. More than 20 AI researchers, including Geoffrey Hinton, Yoshua Bengio, and OpenAI’s Jakub Pachocki, recently warned of a possible "intelligence explosion" in a research paper.
They argued that AI systems could automate all AI research and development within a few years, compressing what would normally be years of progress into just months. That warning describes a possible future; the measures Belfield reported show no current acceleration in research driven by AI itself.
For now, the evidence points to expanding assistance in concrete coding work alongside persistent human oversight and limits on complex autonomous tasks. Tracking both capability gains and the kinds of work AI can complete independently will help show whether that balance changes.