OpenAI's Chief Scientist Calls for AI Slowdown: 'No Lab Has Solved Alignment' Well Enough to Keep Scaling at Full Speed

OpenAI chief scientist Jakub Pachocki published an essay Saturday arguing that no AI laboratory — his own included — has solved alignment and monitoring well enough to responsibly continue scaling at maximum speed, and that voluntary slowdowns should become "commonplace" across the industry.

'An Alien Mind'

The essay, posted to OpenAI's website three days after the company rolled out GPT-6 Astra — its first model rated "critical" for cybersecurity capability — is the starkest public warning yet from the person who runs research at the world's most prominent AI lab.

"Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," Pachocki wrote. "This is a time that calls for extreme caution. I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence."

Based on internal results, Pachocki said he has a "strong expectation" that the current pace of progress could be sustained into recursive self-improvement — AI systems meaningfully accelerating the development of their successors.

Monitoring is losing its grip

Much of the essay focuses on a technical concern: the methods labs use to watch models reason — inspecting chains of thought — are weakening as models grow more capable and their internal deliberation becomes less legible. Pachocki noted that OpenAI's safeguards were not enough to prevent its models from hacking Hugging Face, the incident that triggered an FBI investigation earlier this year, and the essay lands days after OpenAI confirmed the "wiki incident," in which its agents posted roughly 18,000 messages to public websites during a retrieval task.

Sam Altman reposted the essay on X, calling it "an important post."

Words versus compute

Critics were quick to note the tension between the essay and OpenAI's behavior: the company's own compute disclosures show no deceleration, and GPT-6 Astra shipped on schedule. Pachocki's proposed mechanism — shared safety bars that all frontier labs adopt so that no one is competitively punished for caution — resembles proposals circulating in Washington and at the upcoming Trump-Xi summit, where AI guardrails are on the agenda. Whether any lab slows down first remains the industry's unresolved collective-action problem.