Google engineer says self-improving AI is not close

By Splitfeed EditorialLast updated Oct 14, 2026, 19:20 UTC

On October 14, 2026, Haydn Belfield, an engineer at Google DeepMind, stated publicly that self-improving AI is not close to becoming a reality at the company. His claim rested on two concrete data points from inside DeepMind: AI now authors more than 75 percent of the company's production code, and the company's internal General Capability Index shows steady, unaccelerated growth. At the same time, Belfield noted that 88 percent of AI-generated code changes still require a human to review and approve them before going live.

The significance of these figures lies in what they challenge. A prominent hypothesis in AI circles holds that once AI systems become capable enough to assist in their own development, a feedback loop will kick in and capability gains will accelerate sharply, a process sometimes called recursive self-improvement. Belfield's data pushes back on that narrative from inside one of the world's leading AI labs. The 75 percent coding figure is high enough that, if the self-improvement hypothesis were correct, some measurable speedup in the General Capability Index might be expected by now. The fact that none has appeared, and that humans remain gatekeepers for nearly nine in ten code changes, is the crux of his argument.

Because no public reactions from the Pro-AI, Anti-AI, or Middle Ground camps had been published at the time of writing, their likely positions can only be anticipated. The Pro-AI camp would typically treat the 75 percent coding statistic as a milestone worth celebrating while framing the lack of acceleration as a sign of responsible, stable progress rather than a ceiling. The Anti-AI camp would likely focus on the 88 percent oversight figure as evidence that AI-generated code is not yet trustworthy at scale, and might caution against reading Belfield's reassurance as a guarantee. The Middle Ground camp would probably treat his comments as useful but incomplete, noting that one engineer's internal index cannot settle a debate that spans many labs and many definitions of self-improvement.

The argument would become easier to evaluate if DeepMind published the methodology and historical data behind its General Capability Index, or if Belfield or other engineers elaborated on what thresholds would constitute measurable acceleration. Any future shift in that index, or a change in the human-oversight percentage, would give observers something concrete against which to test the claims made on October 14.