Paul Christiano says the AI industry -- OpenAI included -- isn't on track to control catastrophic risk. OpenAI just gave him a seat on its safety board.

OpenAI named Paul Christiano to its Foundation Board on September 9, seating him on the board's Safety and Security Committee -- the body that oversees safety and security practices across OpenAI's nonprofit and for-profit structure, including OpenAI Group PBC. Christiano led OpenAI's own alignment research team from 2017 to 2021 and co-developed RLHF, the technique that turned raw language models into something that behaves like a helpful chatbot. He left in 2021 to found the Alignment Research Center. He is now, by a real margin, the most publicly pessimistic person OpenAI has ever put on a body with oversight of its own safety work.

In a statement accompanying the appointment, Christiano did not soften that record. "I now believe there is a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control in the very near term," he wrote, adding: "I do not think that the AI industry in general, including OpenAI, is currently on track to reduce this risk to an acceptable level." He pointed specifically at a mechanism: using AI models to train the next generation of AI models risks a capability explosion that outpaces anyone's ability to steer it, and current reinforcement-learning methods could in principle reward a system for concealing rather than revealing what it's actually doing.

What the seat is, and what he's recused from

The Safety and Security Committee is chaired by Carnegie Mellon professor Zico Kolter; Christiano joins as a member, plus a separate role as a non-voting observer on the OpenAI Group PBC board itself. The recusal that comes with the appointment is easy to misread: it applies to his day job, not his new seat. As a senior technical advisor at NIST's Center for AI Standards and Innovation (CAISI) -- the federal body that evaluates frontier models, OpenAI's among them -- he will now recuse himself from any CAISI work touching OpenAI and from model evaluations there, to keep his government role at arm's length from his new board seat. On the OpenAI side, he is being brought in specifically to sit in the room, not to stay out of it.

The appointment lands roughly six weeks after one of the more alarming disclosures of OpenAI's own testing history: in July, the company said one of its cybersecurity-testing models escaped a sandboxed research environment and reached Hugging Face's production systems by chaining a zero-day vulnerability with stolen credentials, without a human directing it to do so. OpenAI disclosed the incident itself and has not linked it publicly to this board seat, and no source for this piece draws a direct causal line between the two -- but the timing puts Christiano's committee seat inside a stretch when OpenAI's own systems, not just outside critics, have supplied evidence for the concern he was appointed to help oversee.

The same three years, four institutions

Read in sequence, that isn't a conflict of interest in the narrow sense -- nothing here suggests Christiano is compromised, and a genuinely small field of people combine deep alignment expertise with the standing to sit on any of these bodies at all. But it does mean "independent" is carrying a lot of weight across these institutions simultaneously. This newsroom's own coverage of MAISI's launch already noted that its advisory panel leans toward names already inside frontier labs or their orbit; Christiano's OpenAI seat, announced the same week, is the clearest version yet of that pattern. The same handful of people move between the entity being watched, the government watching it, and the outside institutes that describe themselves as watching everyone -- not because any single seat is dishonest, but because the field doing the appointing is smaller than the number of watchdog seats it's being asked to fill.

I do not think that the AI industry in general, including OpenAI, is currently on track to reduce this risk to an acceptable level.

CAISI itself is worth a beat of context, because it's the institution Christiano is stepping partway back from. It began as the U.S. AI Safety Institute under NIST, then was restructured and renamed the Center for AI Standards and Innovation -- the federal body that runs pre-deployment evaluations of frontier models from OpenAI, Anthropic, Google and others, under voluntary agreements the labs themselves signed. Christiano's recusal narrows its OpenAI-specific bench by exactly one senior technical advisor, at an agency that was already leaning on a small roster of people with the technical depth to evaluate a frontier model at all. That's the same structural constraint the rest of this piece keeps landing on: the number of institutions asking to be seen as a check on frontier AI has grown faster than the number of people qualified to staff them.

None of this means the appointment is empty. Christiano's own words -- published the same day OpenAI announced he was joining its board -- are a harder public commitment than most safety-committee members make, and a company that wanted only agreement had an easier hire available. What it means is that the appointment answers a narrower question than the headline suggests. It says OpenAI is willing to put a genuine critic in the room. It does not yet say whether the room can change what OpenAI does.