Paul Christiano joins OpenAI's Foundation board, and gets no vote at the company

OpenAI has hired a man who thinks we may lose control of AI soon, and given him no vote in the place where the models are made.

Paul Christiano joined the OpenAI Foundation Board on 9 September 2026, along with a seat on the Foundation's Safety and Security Committee, the body that governs safety practices across all of OpenAI including the operating company. He also becomes a non-voting observer on the OpenAI Group PBC board.

Read OpenAI's account of why he came, and then read his.

Throughout Paul's career researching AI alignment, he brings a perspective that takes seriously the possibility that advanced AI could pose catastrophic risks and that developers must prepare before systems reach more dangerous levels of capability.

Here is Christiano, quoted the same afternoon:

Based on the recent trajectory of capabilities and the continued difficulty of alignment, 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.

The distance between those two passages is the whole story. One says a researcher takes catastrophic risk seriously as a category. The other says he has updated, recently, on timing. We read OpenAI's announcement in full. The phrase "near term" does not appear in it.

Credentials are not the question

Christiano led alignment research at OpenAI from 2017 to 2021 and did foundational work on reinforcement learning from human feedback — which is to say he helped build the technique that made this generation of assistants usable at all. He founded the Alignment Research Center, a nonprofit focused, in OpenAI's own words, on aligning advanced AI systems with human interests. Then he went to government, and he is now a senior technical adviser at CAISI inside NIST, where he has spent two administrations evaluating frontier models for national security implications. Kevin Roose called him one of the roughly three most credible AI safety experts in the world, which is not a stretch.

So look at the governance instead.

He votes on the Foundation, the charity that holds control of and a large equity stake in the company. He observes, without a vote, at the public benefit corporation — the entity that trains models, ships them and prices them. The person brought in because he thinks loss of control might be near-term has a vote in the place that grants money, and none in the place that builds the thing.

What do you call oversight with no vote?

The fair answer, and its problem

OpenAI's answer is that the Safety and Security Committee is the lever, and its remit does reach the operating company. That is true on paper and we would not dismiss it. But committee remit without a vote is influence, and influence is measurable only in the things it stops. Zico Kolter has chaired that committee since 2024, and nobody outside OpenAI can name a launch it delayed.

This is a real appointment with a real constraint attached, and Christiano probably understands both halves better than his critics will assume. He has spent a decade arguing that you get more safety inside the room with imperfect authority than outside it with a blog. His joining statement reads like somebody who priced the trade and took it anyway — the phrase "I now believe" is doing a lot of quiet work.

The test is simple and it has a clock on it. If within twelve months the Safety and Security Committee discloses a specific decision it changed, a launch held or a capability withheld or a deployment narrowed, the seat was worth taking. If the only public trace of Paul Christiano at OpenAI a year from now is this announcement, we will have learned something about the seat. So will he.