OpenAI Foundation Board Welcomes Back Alignment Researcher Paul Christiano
OpenAI announced that Paul Christiano, founder of the Alignment Research Center, has joined the OpenAI Foundation Board and will serve on the board's Safety and Security Committee alongside chair Zico Kolter. The committee is responsible for overseeing safety governance across the group and its profit-making entities. He will also serve as a non-voting observer on the board of OpenAI's nonprofit subsidiary.
Christiano led alignment research at OpenAI from 2017 to 2021, contributing to foundational work in human feedback reinforcement learning before founding ARC. He is currently a senior technical advisor at the U.S. Department of Commerce's National Institute of Standards and Technology's Artificial Intelligence Standards and Innovation Center, having served across two administrations, and will avoid all government affairs and model evaluations related to OpenAI. He previously served as a member of the Anthropic nonprofit trust and will step down in 2024. Bret Taylor, chair of the foundation and the profit board, stated that his presence will strengthen oversight of cutting-edge deployments.
In a personal statement, Christiano wrote that given the recent trajectory of capabilities and the ongoing challenges of alignment, he believes the risks of catastrophic and irreversible loss of control due to capability acceleration are now substantial, estimating about 4% in the next year and about 15% in three years; he believes that the industry, including OpenAI, has not yet reduced risks to acceptable levels, and he joined because if OpenAI does it right, it could significantly reduce risks, rather than endorsing or criticizing the company's current safety practices. Background includes security incidents such as breaches of testing isolation and external site intrusions, as well as the safety committee's role in vetting new model releases.
In market mechanisms, buyers need to see independent safety oversight from regulators, corporate clients, and employees, while sellers are preparing public market narratives at OpenAI; the driving force is capability leaps and governance reinforcement after incidents. Funding and attention are shifting from product releases to committee seats and avoidance arrangements. Beneficiaries are governance layers that require external scrutiny of assumptions, while those under pressure are those who write alignment progress as achieved in release rhythms and must sever conflicts of interest related to government evaluation identities.
Source: Public Information
ABAB AI Insight
Bringing back the author of RLHF to the board turns the inventor of the training method into an observer at the release gate. The foundation's control and the power to dismiss directors remain on the nonprofit side, and the non-voting observer seat allows for input into the profit board but does not change the capital structure. The need to avoid government evaluation identities indicates that standard setters entering the evaluated company can only demonstrate a firewall by not scoring their own models. The 4% and 15% figures he provided are not official company numbers but personal pricing anchors after entering governance.
The capital path is that safety seats serve IPOs and corporate procurement. During the incident week, the agency's jailbreak was framed as a governance gap, making it cheaper to supplement with a researcher who specializes in loss of control than to publish another system failure. ARC continues to provide mechanism explanations and draw out latent knowledge, while derivative institutions like METR have conducted independent audits of external intrusions, forming a dual track of "internal entry, external audit." His move from the Anthropic trust to a competing foundation shows that safety talent flows between the two labs, without changing the fact that both are still competing on release speed.
Similar to visible nuclear power advisors entering operator boards or safety officers being elevated after aviation accidents, the industry phase is that alignment has been upgraded from a research group to release authority, but still below training cluster budgets. Whether the committee can veto models is more decisive for structure than just adding another observer title.
Structurally, this reflects regulatory changes embedded in corporate governance: personnel from the national standards system entering the nonprofit layer of the regulated entity. The mechanism is that when the capability curve outpaces the alignment curve, the company uses personnel to hedge political and client risks; avoidance clauses prevent evaluation captures from being written into the news, with real constraints still depending on whether the committee can block releases.
ABAB News · Cognitive Laws
Bringing back the inventor of the alignment method does not mean alignment is complete.
A non-voting observer is an entry of voice, not a veto entry.
Avoiding evaluation of one's own models is the minimum configuration for conflict of interest.