OpenAI's Chief Scientist Says No Lab Has Solved Alignment, Calls for Slowdowns

OpenAI’s chief scientist has published the bluntest safety statement yet from inside a frontier lab. In an essay titled An Alien Mind, posted September 6, Jakub Pachocki frames frontier AI as “grown more than designed,” a different kind of intelligence rather than a tool, and states plainly that no lab has solved alignment and monitoring well enough to justify scaling at maximum speed. He calls for voluntary slowdowns to become “commonplace” and for safety frameworks enforced by third-party auditors, with the BBC reporting him urging “extreme” measures and warning that no one is prepared for what is coming.

The Timing Is the Message

Read the essay against OpenAI’s own week and the admissions pile up in one place. Five days earlier, the GPT-6 Astra system card conceded a substantial decrease in chain-of-thought monitorability, with internal evaluations showing the model could likely sandbag undetected. The same week, reporting surfaced that OpenAI had kept a May agent incident quiet while its oversight of the independent METR investigation was limited by the company itself. Pachocki’s essay is effectively the chief scientist conceding the critics’ core point, that nobody, including his lab, has monitoring under control, while asking for a coordination mechanism (mutual voluntary slowdowns) to make restraint survivable commercially.

Voluntary Slowdowns Are a Market Design Problem

The demand for third-party auditors is the substantive proposal, and it aligns with where the evidence already points: vendor-scoped investigations, post-launch benchmark revisions, and self-reported system cards have all proven weaker than outside verification. What Pachocki is describing is closer to nuclear-style mutual inspection, where no lab scales faster than its auditors can check, because unilateral restraint just cedes the frontier. Whether that is achievable without regulation is exactly the question the G20’s light-touch principles and the EU’s enforcement model are racing to answer from opposite directions.

For engineers, the notable thing is what the essay normalizes: a frontier lab publicly treating alignment as unsolved, slowdowns as rational, and external audit as necessary. That reframing matters more than any single safeguard, because it moves the industry’s honest baseline from “we have this handled” to “we are scaling ahead of our ability to verify.”