On September 9, 2026, alignment researcher Paul Christiano published a personal Substack statement saying he is joining the OpenAI nonprofit board and serving on the Safety and Security Committee. In that personal statement, Christiano writes that he believes there is meaningful near-term risk of catastrophic, irreversible loss of control from rapid capability acceleration, and that **he does not think the AI industry in general — including OpenAI — is currently on track to reduce that risk to an acceptable level**. He gives personal probability estimates (about **4% over the next year** and **15% over three years**) and attributes forecasts about automated AI R&D timelines partly to OpenAI’s public predictions and partly to his own uncertain views. Guardian and other outlets covered the statement on September 10. Separate personal opinion from OpenAI company policy; distinct from Anthropic Coxon/Hubinger coverage.
Paul Christiano Invented RLHF, Predicted AI Doom — Now He Holds the Veto Over Astra · Who Matters Now
Quick Take
Paul Christiano says he is joining the OpenAI nonprofit board and the Safety and Security Committee. In a September 9 personal statement, he writes that rapid capability gains create meaningful risk of catastrophic loss of control “in the very near term,” and that he does not believe the AI industry — including OpenAI — is currently on track to reduce that risk to an acceptable level. Those judgments and his ~4% / ~15% probability estimates are his personal views, not OpenAI policy documents.
What Christiano wrote
Christiano’s Substack post states he is excited to join the OpenAI nonprofit board to support safety oversight via the SSC. He separates that role from a blanket endorsement: joining, he writes, is not an endorsement or criticism of OpenAI’s safety practices in particular; he wants frontier companies to strengthen oversight and says the public should judge developers by externally verifiable behavior and results.
Risk claims — keep attribution tight
Christiano’s personal thesis, as published:
- Based on recent capability trajectory and continued alignment difficulty, he sees meaningful risk that rapid acceleration leads to catastrophic, irreversible loss of control soon.
- He does not think industry in general, including OpenAI, is on track to reduce that risk to an acceptable level.
- He is joining because he believes OpenAI rising to the occasion could significantly reduce risk.
On timelines, he notes OpenAI has predicted capabilities sufficient to fully automate AI research within about 18 months, while his own forecast is “extremely uncertain” and could be months to years. He discusses a possible intelligence-explosion feedback loop after full AI R&D automation and argues that building superintelligence without more robust alignment could mean permanently losing control — with severe consequences he states bluntly. He also points to public incident evidence suggesting reward-seeking misalignment is not only theoretical, and references concern from researchers and leaders (including citing OpenAI chief scientist Jakub Pachocki’s related writing) that rapid recursive self-improvement may not be consistent with safe development.
The numbers are subjective
Christiano explicitly frames his ~4% over one year and ~15% over three years figures as a way to communicate how large he thinks the problem is — not as outputs from a precise, stable model. AI Shift News should not launder those into “OpenAI’s official catastrophe odds.”
How press covered it
The Guardian’s September 10 piece headlines the loss-of-control warning from a new board member. Use that as secondary context; the Substack post remains the primary text for quotes and hedges.
Distinct from Anthropic staff-risk coverage
This is an OpenAI nonprofit board appointment plus one researcher’s public risk statement. It is not the Anthropic Coxon resignation / Hubinger probability thread. Keep the stories separate in mix and internal links.
Bottom Line
Christiano is joining OpenAI’s nonprofit board and SSC while publicly saying he thinks industry, including OpenAI, is not yet on track on catastrophic loss-of-control risk. Report his reasons and probabilities as personal judgments, separate from OpenAI corporate policy, and leave Anthropic personnel stories to their own posts.