Google: Self-Improving AI Is Not Close, Engineer Says

By Rebecca Stone

Google engineer Haydn Belfield says self-improving AI is not anywhere close, and that DeepMind has measured no acceleration of AI research driven by AI itself.

The claim cuts against one of the most consequential bets in the industry: that AI systems will soon improve their own capabilities and compress the timeline to more powerful models. Belfield, speaking from inside Google's AI organization, reports the opposite pattern at the company best positioned to observe it.

What did Google actually measure?

Belfield says there is no measurable acceleration of AI research driven by AI at DeepMind. The company tracks this question directly. Its in-house General Capability Index shows steady growth in system capabilities, with no sudden speedup.

That index matters because it is Google's own instrument for detecting exactly the kind of nonlinear takeoff that self-improvement scenarios require. Steady growth, recorded by the company building the systems, is the strongest available counter-evidence to claims of an imminent feedback loop.

How much code does AI write — and who checks it?

The numbers Belfield cites come from Google's software engineering operations, the largest codebase in commercial software.

• AI now writes more than 75 percent of production code.
• 88 percent of those changes still need human oversight.

Read together, the two figures define the current boundary of automation. Machines generate the majority of code, but humans remain in the loop for the overwhelming majority of changes. Generation has scaled. Verification has not.

Why does this matter?

The stakes run in two directions. For AI labs and their investors, self-improvement is the mechanism behind the most aggressive forecasts of capability growth and the capital budgets built on them. If research progress remains linear — driven by human researchers, compute and data — timelines lengthen and valuations face re-examination.

For policy, the same question anchors debates over regulation and safety timelines. Governments preparing for rapid capability jumps are working from assumptions that Belfield's account does not support, at least at Google.

What is the state of the takeoff debate?

The debate over recursive self-improvement has divided researchers for years. One camp argues that as AI takes over more of the research workload — writing code, running experiments, drafting papers — progress compounds. The other camp holds that the hard parts of research, namely judgment, verification and taste, resist automation.

Google's internal data, as reported by Belfield, sits squarely with the second camp. The 88 percent human-oversight figure shows where the bottleneck sits: not in producing code, but in checking it.

What happens next?

Belfield's assessment is that self-improving AI isn't anywhere close. The General Capability Index will continue to be the metric to watch: a visible inflection there would be the first hard signal that the feedback loop has started.