Google engineer Haydn Belfield stated that although the degree of AI automation in model R&D continues to rise, there is currently no measurable evidence that AI self-driven research is accelerating.
Belfield pointed out that AI currently writes over 75% of production code, but about 88% of code changes still require human oversight or review, and every change must undergo human review before submission. The proportion of code changes handled entirely by AI has risen from 7.5% to 11.9%. In 1389 internal Argon agent run tests, the higher the task complexity, the lower the AI's success rate without human intervention.
A survey of 12 senior DeepMind researchers also showed that AI has a high degree of automation in well-defined tasks with verifiable outcomes, but still relies on humans for creative research, experiment design, training anomaly diagnosis, and long-term planning.
On the RE-Bench benchmark, which contains 74 real machine learning engineering tasks, Argon's task-solving rate was about 50%, higher than Gemini 3.7 Flash's 27%, but still short of achieving full automation.
To track changes in AI capabilities, the research team built a General Capability Index (GCI) covering 123 models and 92 benchmarks. The data shows that Gemini model capabilities are growing steadily at a rate of about 0. 1 GCI points per month, with no statistically significant sign of acceleration yet. The team believes that Gemini 4 Argon's performance, slightly above the growth trend line, mainly comes from increased compute and organizational investment, rather than AI self-driven R&D acceleration.
Currently, the research team believes the risks from AI-driven research automation remain relatively limited, but they are continuously improving their monitoring systems. Meanwhile, more than 20 researchers, including Geoffrey Hinton, Yoshua Bengio, and OpenAI chief scientist Jakub Pachocki, have warned that AI may achieve a high degree of automation of R&D work in the coming years, compressing years of progress into months. These differing judgments highlight that considerable uncertainty remains regarding AI's self-improvement capabilities and their potential risks.