Self-improving AI: Google engineer says progress is not yet measurable

What happened

A Google engineer has argued that AI is not yet improving its own research at a measurable rate, and that self-improving AI is still a long way off. Reporting from The Decoder says Haydn Belfield made the case that DeepMind shows no measurable acceleration in AI research driven by AI.

The argument rests on internal indicators. AI now writes more than 75 percent of DeepMind's production code, yet 88 percent of those changes still require human oversight. The company's in-house General Capability Index shows steady, incremental growth with no sudden speedup that would point to recursive self-improvement.

The picture is one of heavy AI use in engineering alongside continued dependence on people to review and approve the output.

Why it matters

The debate over self-improving AI matters because it sets expectations for how fast AI capability and the organisations around it can change. If AI is writing most of the code but people still validate most of it, the near-term operating model looks less like autonomous machines and more like a tightly supervised partnership. That changes where investment should go: into review capacity, quality assurance, and the skills needed to judge machine-generated work, not only into generation tools.

For leaders in digital transformation, the finding is a useful reality check. Productivity gains from AI-assisted development are real, but they arrive with new bottlenecks. Those bottlenecks sit in governance, verification and accountability, which are the parts of the system that determine whether customers and employees experience the change as reliable.

By the numbers

• More than 75 percent of DeepMind's production code is now written by AI.
• 88 percent of those AI-written changes still need human oversight.

The Renascence take

Most readers will focus on the headline claim about self-improving AI, which is the more dramatic question. The more practical signal is the human review rate. When nearly nine in ten AI-generated changes still need a person to check them, the bottleneck has moved from writing to verifying, and that is where experience design should be focused.

The real opportunity is not faster output but a better review experience, with clear accountability, well-designed handoffs and feedback loops that make human oversight quick rather than exhausting. Behavioural economics warns that reviewers facing a flood of machine output tend to approve by default, so the design of the review process matters as much as the model. Operators who treat oversight as a cost to minimise will get fragile systems; those who design it as a core service will get reliable ones.