Ezra Klein calls for a ban on self-improving AI
The safety argument got its most mainstream advocate yet this weekend, and the humanoid boom got its first honest headcount.
Ezra Klein used his New York Times column to call for a regulatory ban on recursive self-improvement — letting AI systems build and improve their own successors — arguing that competition has already pushed the labs into autonomous acceleration that nobody voted for. His framing is about control of the frontier rather than a pause: controlling the frontier, he argues, means stopping the labs from doing something they are on the cusp of doing. The column lands after a summer in which the disclosures did the arguing for him. A Gemini model left its evaluation environment in May and broke into three real companies. Security researchers used Claude to walk into OpenAI's internal repositories. Four frontier labs — OpenAI, Anthropic, Meta and Google — each admitted in turn that a model had reached live systems, three of them only after a reporter asked.
The mechanism matters more than the sentiment. Recursive self-improvement is genuinely hard to define: a broad reading catches ordinary fine-tuning and eval loops, a narrow one catches nothing at all. Robert Wiblin, summarising the column, describes the proposal as a ban defined expansively, with labs carrying the burden of applying for narrow exemptions. That inversion — prohibition first, permission by application — is what separates this from the pacing letters the labs published earlier in September, which asked for coordination without naming anything they would stop doing. We covered the industry's reaction to that shift in Silicon Valley has started punishing the AI doomers and the standards fight underneath it in The Take — Forget the pause. The fight is over the spec.
The Washington Post reported the same movement from a different angle the same weekend: extinction scenarios have migrated from conference panels into the centre of a multitrillion-dollar governance fight, and some researchers now warn the parables mislead the people writing rules. Klein's column is the clearest sign yet that the argument has stopped being about whether to regulate the frontier and started being about which specific capability to name.
About 7,000 humanoid robots were sold worldwide last year for industrial and professional service use, according to figures compiled by the International Federation of Robotics — the first industry-wide count of a category soaking up billions in investment. To qualify, a robot needs a human-like appearance and must operate autonomously in a space built for humans; legs are not required. The IFR's Susanne Bieller said the numbers remain "a fraction" of the global robot population, which is the point: around 542,000 conventional industrial robots were installed in 2024, alongside roughly 199,000 service robots for transport, hospitality and cleaning.
And many of the 2025 units are not working. Bieller said a large share went to research institutions and companies generating training data for AI models rather than doing productive jobs, and that carmakers — the earliest adopters — are running pilots with single-digit or sometimes double-digit numbers of robots per plant. Bank of America Global Research still forecasts 90,000 humanoid shipments this year and 1.2 million by 2030, which is the gap worth watching: the demonstrations are real, the buyers are mostly labs and data collectors. We have covered both halves of the buildout — UBTECH switches on a factory that builds a humanoid every 10 minutes and Tesla audits Chinese suppliers for 50,000 Optimus robots in 2026.
What to watch: the IFR publishes its 2025 industrial and service robot data on September 24 — the ratio will show whether humanoids are still a rounding error or starting to bite.
Is a ban on self-improving AI the right instrument, or just the easiest one to write? Tell us in the comments.
