Google says self-improving AI isn't anywhere close

Google DeepMind engineer Haydn Belfield pushed back on claims that AI systems are already meaningfully accelerating AI research. He noted that while AI-generated code now accounts for more than 75 percent of production code at the company, roughly 88 percent of those changes still require human review before they can be merged.

The remarks land amid growing hype around self-improving AI and recursive automation. Belfield's comments suggest that current systems remain assistive tools rather than autonomous drivers of research progress, and that human oversight remains a binding constraint on throughput.

Google also referenced an in-house General Capability Index showing steady but incremental gains rather than step-change breakthroughs. If accurate, this tempers expectations for near-term autonomous AI research and underscores that scaling gains may still depend heavily on human-in-the-loop workflows.

For the broader industry, the message is a reality check: even at a frontier lab where AI tooling is deeply embedded in engineering, the loop is not yet closed. Progress appears to be compounding, but the last mile still runs through human judgment.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.