Google engineer says self-improving AI is still far off despite rising code automation

Google DeepMind has seen no measurable acceleration of AI research driven by AI itself, according to engineer Haydn Belfield. The finding pushes back against warnings from more than 20 prominent AI researchers who recently cautioned that an "intelligence explosion" could compress years of progress into months. For executives and product leaders weighing automation investments, the data offers a reality check on what self-improving AI can and cannot do today.

Belfield shared the findings in an October 2026 safety report. AI now writes more than 75 percent of production code at Google, but about 88 percent of code changes still need human oversight or review. Every change is reviewed by a human before submission. The share of code changes fully delegated to AI rose from 7.5 to 11.9 percent - a meaningful increase, but far from autonomous development.

Complexity still demands human judgment

The more complex and time-consuming a task gets, the less likely AI is to handle it on its own. Across 1,389 internal Argon agent runs, the success rate without human input dropped sharply as complexity increased. Belfield said humans remain essential for creative research, novel experiment design, diagnosing subtle training instabilities, and long-term planning.

A survey of twelve senior DeepMind researchers confirmed the pattern. Automation ran highest for well-defined, verifiable tasks and lowest for open-ended scientific ideation. On DeepMind's in-house benchmark RE-Bench, which includes 74 realistic ML engineering tasks, Argon solves about half of them - a big jump from Gemini 3.7 Flash at 27 percent, but still far from the near-perfect performance full automation would require.

No speedup in AI research capability

To track progress systematically, the team built a General Capability Index covering 123 models and 92 benchmarks. Gemini model capabilities are growing at a steady pace of about 0.1 GCI points per month, with no significant change in speed. The team attributes Gemini 4 Argon sitting slightly above the trend line to more compute and organizational investment rather than AI-driven acceleration.

The team currently rates the risk from AI-driven research automation as low but is building measurement infrastructure to track signs of acceleration. This measured stance contrasts sharply with a recent research paper signed by Geoffrey Hinton, Yoshua Bengio, OpenAI's Jakub Pachocki, and others. 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.

Why this matters for executives and product leaders

The gap between AI-assisted coding and autonomous research matters for resource planning. Code generation tools are delivering real productivity gains - 75 percent of production code is a number that demands attention. But the 88 percent human review rate and the steep drop in agent success on complex tasks mean engineering teams still need senior talent who can design experiments, spot subtle failures, and make architectural decisions. Automation investments should target well-defined, verifiable workflows where the success criteria are clear. For teams building internal AI tools, AI Agent Courses can help developers understand where current agents excel and where they break down.

For R&D organizations, the message is clearer: don't budget for self-accelerating AI timelines yet. Google's own capability index shows linear progress, not exponential. The researchers building these systems are themselves building the dashboards to detect if that ever changes. When they see acceleration, they'll say so. They haven't.