Google Says Self-Improving AI Remains Far Off
A Google engineer says there is no measurable acceleration of AI research driven by AI at DeepMind, even as AI writes most production code. Human oversight remains essential, and the company's internal capability index shows steady rather than explosive growth.
Google has pushed back against the idea that artificial intelligence is on the verge of improving itself without human help. Haydn Belfield, an engineer at the company, says there is no measurable acceleration of AI research driven by AI at DeepMind, the tech giant's AI division. The assessment challenges a growing narrative that rapid self-improvement could soon lead to runaway machine intelligence.
Belfield's comments come as AI systems take on a larger share of technical work. At Google, AI now writes more than 75 percent of production code. That figure might suggest machines are close to replacing human engineers, but the reality is more complicated. According to Belfield, 88 percent of those AI-generated changes still require human oversight before they can be accepted. The numbers point to a collaborative workflow rather than an autonomous one, with people reviewing, correcting, and approving the bulk of machine output.
The company's in-house General Capability Index, which tracks progress across a range of AI tasks, shows steady growth with no sudden speedup. If AI were beginning to improve itself in a meaningful way, the index would be expected to show a sharp upward curve. Instead, the data indicates incremental gains, consistent with the kind of progress the field has seen in recent years. Belfield's conclusion is blunt: self-improving AI is not anywhere close.
The debate over self-improving AI has gained prominence as large language models and other systems have become more capable. Some researchers and commentators have warned that once AI can rewrite its own code or design better versions of itself, the pace of change could become uncontrollable. Google's position offers a counterweight to those fears, at least for now. The company is not dismissing the long-term possibility, but it says the evidence does not support claims of an imminent breakthrough.
For the broader technology industry, the findings highlight a persistent gap between automation and autonomy. AI can generate code quickly, but it still struggles with the judgment, context, and accountability that human engineers provide. That gap has implications for how companies deploy AI, how they train workers, and how they manage risk. It also shapes public expectations, which often run ahead of what the technology can actually deliver.
Google's internal metrics are not the only measure of progress, and other labs may report different results. But the company's willingness to share its assessment adds a note of caution to a conversation that has been dominated by bold predictions. As AI systems become more integrated into software development and other fields, the question of whether they can truly improve themselves will remain central to both research and policy. For now, the answer from Google is that the era of self-improving AI has not arrived, and there is no clear sign of it arriving soon.
5
