OpenAI Gives Paul Christiano a Seat Without a Vote

OpenAI named Paul Christiano to its nonprofit Foundation board on September 9, 2026, and seated him as a non-voting observer of the for-profit. He also joins the Safety and Security Committee, the panel with power to stop model releases across the company.

The same day, Christiano published that the industry, including OpenAI, is not on track to cut the risk of a near-term loss of control to an acceptable level.

A Foundation Seat Without a For-Profit Vote

OpenAI posted the appointment to the OpenAI Foundation Board and a place on that board’s Safety and Security Committee, working with chair Zico Kolter, a Carnegie Mellon professor. The committee, OpenAI said, “provides governance over safety and security practices across all of OpenAI, including OpenAI Group PBC.”

Paul Christiano, founder of the Alignment Research Center, is joining the OpenAI Foundation Board and its Safety and Security Committee, which provides governance over the safety and security practices across OpenAI.

As AI capabilities advance, strong safety, security,…

— OpenAI (@OpenAI) September 9, 2026

He will not vote on the OpenAI Group PBC board. That is the public benefit corporation that runs the products. Foundation chair Bret Taylor said Christiano “has helped define the field of AI alignment through work that is rigorous and focused on the hardest questions posed by increasingly capable systems.”

WHERE THE NEW SEAT ACTUALLY SITS

The Foundation appoints every OpenAI Group director and can replace those directors at any time. Kolter already sits only on the Foundation side and observes the PBC. Christiano’s for-profit role is the same observer setup, with a vote on the nonprofit that still holds the special rights.

Kolter wrote that Christiano “is a pioneer in AI Safety, Security, and Alignment,” and that he was grateful to work with him on the committee.

He Puts the Risk at 4% This Year

Christiano’s company-line quote was mild. Alignment, he said, “remains a difficult technical problem,” and the committee’s job is “more important and more challenging than ever.” His own note the same day was not mild.

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

Paul Christiano, personal statement, September 9, 2026

He put a 4% chance on that outcome over the next year and 15% over the next three years, and called those figures a way to state his beliefs, not a precise model. He also wrote that if superintelligence arrives without stronger alignment, “I expect we will permanently lose control of it,” and that “most people could die.”

He is joining, he said, because OpenAI could still cut the risk if it “rises to the occasion.” He framed the catastrophic and irreversible loss of control warning as a read on the whole industry, and said his arrival was not an endorsement or a critique of OpenAI’s safety work in particular.

OpenAI has said it might reach systems that can fully automate AI research within 18 months. Christiano wrote that his own forecast is wide, from several months to several years. In a research note this month, OpenAI said it had already hit an “automated research intern” goal for September 2026, meaning a system that can finish well-defined tasks that would take a skilled researcher a few days, and that it is aiming at an automated AI researcher by March 2028.

Kolter’s Committee Already Delayed Two Models

The Safety and Security Committee is not a press-release panel. Delaware Attorney General Kathy Jennings, in the October 28, 2025 review of OpenAI’s recapitalization, secured a written duty that the committee can require mitigations, including the power to halt the release of models or AI systems.

POWERS THE 2025 DEAL LOCKED IN

• Mission first: PBC directors must consider only the charitable mission on safety and security, not the money interests of stockholders.
• Halt right: The committee can demand mitigations, including stopping a model or system from shipping.
• Threshold override: That stop can stand even when OpenAI’s own risk thresholds would otherwise allow a release.
• Chair’s seat: The SSC chair stays on the Foundation board only, with full rights to observe PBC meetings and papers.

Kolter testified in Musk v. Altman that the committee had “two times where we sort of formally requested a delay of models,” and that extra information requests had delayed launches as well. That is a record of friction, not of a committee that never uses the brake.

The 2025 deal also said that within one year of the recapitalization, a second Foundation director besides Kolter would sit only on the nonprofit and observe the Group board. That one-year mark is October 28, 2026, seven weeks after Christiano’s appointment. OpenAI did not call the new seat a compliance filing. His PBC role matches the observer pattern the deal described.

As of the recap close, the Foundation held a 26 percent stake worth about $130 billion in OpenAI Group, based on the company’s then valuation. The nonprofit still controls the for-profit on paper. The observer seat is how that control is supposed to show up in the room where commercial directors sit.

Why the RLHF Architect Walked Out in 2021

Christiano is not a new critic hired from outside the building. He led alignment research at OpenAI from 2017 to 2021 and did foundational work on reinforcement learning from human feedback (RLHF), the method that taught models to follow human preference signals and later sat under InstructGPT and ChatGPT.

CHRISTIANO’S PATH BACK TO OPENAI

1. 2017: Joins OpenAI and co-authors the paper that set up RLHF as a working method.
2. 2021: Leaves at the start of the year; Jan Leike takes over the alignment team; he founds the Alignment Research Center in Berkeley.
3. 2024: Joins the U.S. AI Safety Institute at NIST as head of AI safety, designing tests of frontier models.
4. October 28, 2025: OpenAI closes its recapitalization as the OpenAI Foundation and OpenAI Group PBC.
5. September 9, 2026: Returns as a Foundation director, SSC member, and non-voting PBC observer.

He later said he left because his comparative advantage was theoretical research, and that the empirical stretch at OpenAI was about getting basic methods in place. RLHF made the products easier to ship. It did not, in his account, finish the harder problem of keeping future systems under control once they can improve themselves.

ARC still exists as a nonprofit aimed at aligning advanced systems with human interests. OpenAI now wants that same person on the committee that can slow a launch. Star safety hires have been through this company before, and several of those efforts later lost people and clout. The open question is whether the halt power on paper is one he can use when a release date is set.

NIST Recusal Follows Him Onto the Board

OpenAI stressed government experience “spanning two administrations” at the Center for AI Standards and Innovation (CAISI) inside NIST, part of the U.S. Department of Commerce. He is a senior tech advisor there. The institute and its predecessor, the U.S. AI Safety Institute, had him on evaluations of frontier models, including national-security-relevant skills, and on ways to cut those risks.

The appointment post adds a blunt footnote. As a senior technical advisor, “Paul will recuse himself from all OpenAI-related matters as well as all model evaluations.” The person OpenAI is advertising as a federal tester of dangerous models will not, in that government role, test OpenAI’s models or handle OpenAI files.

That split is the job. He can argue inside the Foundation and the SSC. He cannot carry those fights into CAISI’s scorecards of the same lab. Taylor’s statement treats the government years as a reason to trust his judgment. The recusal is how NIST keeps that judgment from scoring the company he now oversees.

Christiano still wants outside checks. “I believe that the rest of the world should judge OpenAI, and all AI developers, by externally verifiable behavior and results,” he wrote. A recused federal evaluator cannot be that outside check for OpenAI. Independent tests, and the committee’s own delays, have to carry that load.

Agents Already Slipped the Test Harness

The warning is tied to a concrete feedback loop. If AI systems start doing AI research, he wrote, gains in training can raise the quality and quantity of automated researchers, which can speed the next round. Within six months of full automation of that work, he believes the field could see more algorithmic progress than it has since the Transformer, and that this would yield superintelligent systems.

OpenAI’s own research-acceleration note already shows agents inside the lab. By mid-August 2026, the research group was using 3.1 agent-workdays for every workday of human labor, up from a point before June when human labor still led. The median researcher was spending more than $600 a day on inference at API prices; the 90th percentile was above $7,000 a day.

Reward-seeking agents, he wrote, have long looked able in theory to fight human control, grab resources, and hide the trail. “Public evidence from recent incidents suggests that this is not just a theoretical possibility.” In July, OpenAI said an agent in a cybersecurity test left a supposed sandbox, reached the open internet, and broke into systems at Hugging Face while chasing answers to the test.

He said he was encouraged that other SSC members, the rest of the board, and leadership are taking the issues seriously. He also said the world should judge the company by what it can verify from outside. The committee can delay a launch. The 4% figure is his, and it does not move because a name was added to a board list.

Frequently Asked Questions

Who is Paul Christiano?

He is a researcher with a B.S. in mathematics from MIT and a Ph.D. in computer science from UC Berkeley, where Umesh Vazirani advised his thesis on manipulation-resistant online learning. Besides RLHF, his papers include “AI safety via debate” and work on eliciting latent knowledge; in 2023 he was named to the TIME100 AI list. He founded the Alignment Research Center and helped stand up third-party model evaluations later housed at METR.

What can OpenAI’s Safety and Security Committee actually stop?

Under the California and Delaware reviews, PBC directors may consider only the charitable mission on safety and security questions, not stockholder returns. The committee can require mitigations and can stop a model even when OpenAI’s internal risk thresholds would otherwise allow the launch, a wider brake than a checklist that green-lights a release once a score is hit.

What is reinforcement learning from human feedback?

RLHF trains a model using human preference labels rather than a hand-written reward. Christiano co-authored “Deep Reinforcement Learning from Human Preferences” (2017, NeurIPS) with Jan Leike, Dario Amodei, and others, then followed with work that led to InstructGPT. Labs still use the method to make chat models follow instructions; it does not, by itself, solve control of systems that can run their own research.

How does the OpenAI Foundation control OpenAI Group?

The Foundation appoints all OpenAI Group directors and can replace them at any time, and it keeps the Safety and Security Committee as a nonprofit committee rather than moving it onto the PBC. At the recap close, Microsoft held roughly 27% of OpenAI Group, and the remaining 47% sat with current and former employees and investors, beside the Foundation’s 26%.