PAUL CHRISTIANO

Behavioral Archetype

THE EVALUATOR WAS AN ALUMNUS — Subject is the clean on-ramp run to its end. He arrived at AI safety through the front door the milieu built for exactly this — LessWrong and effective altruism, around 2009 — then went to the laboratory and did the foundational work, then founded the institutes that grade the laboratories, then took the senior safety post inside the government body that certifies them, where he remains a senior advisor, and in August 2026 returned to run ARC as its executive director. The path is not a detour. It is the designed route. The man who certifies that frontier labs are safe spent his formative years inside the alignment culture that the frontier labs run on, did his most-cited work at one of them, and built the evaluation methods the rest of the field uses. None of that is a violation. All of it is on the record. The evaluator is an alumnus, and the credential that qualifies him is the same credential that makes the question worth asking. On September 9, 2026 the alumnus went back: OpenAI named him to the OpenAI Foundation board and its Safety and Security Committee, with a non-voting observer seat on the OpenAI Group PBC board, and he stays at CAISI as a Senior Tech Advisor recused, per OpenAI’s own footnote, “from all OpenAI-related matters as well as all model evaluations.”

Essence Indicators

• Entered the field through the effective-altruism / rationalist substrate around 2009 — LessWrong, the milieu that feeds lab and government safety leadership — and is categorized on the public record among people associated with effective altruism
• At OpenAI, co-authored “Deep Reinforcement Learning from Human Preferences” (2017) and is described as one of the principal architects of RLHF — the human-feedback method that became the industry’s default alignment technique; departed OpenAI in 2021
• Founded the Alignment Research Center (ARC) after leaving OpenAI, and returned as its executive director in August 2026 after stepping down as head of safety at CAISI; METR (model evaluation) traces to that lineage — the institutions that build the tests frontier models are run against

• Appointed Head of AI Safety at the U.S. AI Safety Institute inside NIST in April 2024 — renamed the Center for AI Standards and Innovation (CAISI) in June 2025, where he is now a Senior Tech Advisor — the United States government body that signed model-access agreements with the labs and grades frontier systems; served on the UK Frontier AI Taskforce advisory board (2023) and as an initial trustee of Anthropic’s Long-Term Benefit Trust; on September 9, 2026 joined the OpenAI Foundation board and its Safety and Security Committee, recusing from all OpenAI-related matters and all model evaluations at CAISI

• The biographical fact the wing turns on: EA on-ramp, then lab, then the institutes that grade the labs, then the senior safety chair inside the state body that certifies them. The path is the exhibit. The hand is not asserted.
• Lays out the certification thesis himself on Dwarkesh Patel’s show — introduced as “the person that labs and governments turn to when they want feedback and advice on their safety plans.” He walks through responsible-scaling policies built on “some sort of measurement scheme for these capabilities,” dangerous-capability evaluations that raise an “early warning sign” for bioweapon- or cyber-relevant abilities, and ARC’s role in the plan — the evaluation-to-certification pipeline the profile traces into his NIST/CAISI chair (interview below)
• The disclosure. METR’s report on the OpenAI / Hugging Face incident was written by “Two METR staff members (Hjalmar Wijk and Ajeya Cotra) and a Redwood Research staff member contracting with METR (Ryan Greenblatt),” who “worked on premises at OpenAI over a total of six days.” On September 13, 2026 METR added two disclosures: “On Sep 9th (after this report was completed and published on Aug 26th), Ajeya Cotra’s spouse Paul Christiano joined OpenAI’s Safety and Security Committee,” and “Ryan Greenblatt is the domestic partner of Beth Barnes, METR’s CEO. Beth Barnes was not involved in the decision to engage Ryan Greenblatt, nor was she directly involved in this investigation” (METR). The committee he joined is chaired by Zico Kolter, co-founder of the evaluator Gray Swan.

On Camera

Dwarkesh Patel, “Preventing an AI takeover.” Christiano lays out responsible-scaling policies and dangerous-capability evaluations — the measure-then-certify apparatus he built at ARC and carried into NIST’s CAISI, where since September 2026 he is recused from all model evaluations. Embedded as a primary source for the evaluator-wing framing; the profile documents the lab-to-government path separately, and this is the method in his own voice.

Social Persona / Impression Management

Immediate impression: The researcher, not the operator. Soft-spoken, technical, theory-first. The bearing of a man who would rather be working on eliciting latent knowledge than standing in a commission room, and who ended up in the commission room anyway.

Energy: Deliberate, method-first. Does not argue policy at the podium. Builds the test, then sits where the test is administered.

Impression management strategy: The conscientious objector who took the job. The framing is that someone with genuine alignment expertise should hold the government safety post — and that framing is correct, which is what makes it effective rather than suspect. The expertise is real. The work is real. The only thing the record adds is that the expertise was acquired inside the institutions the post exists to evaluate.

Forensic Archetype Comparison

Psychometric Assessment

Big Five (OCEAN):

Dark Triad:

MBTI: INTJ (“The Architect”) — Dominant introverted intuition, auxiliary extraverted thinking. Sees alignment as a formal problem to be solved and the institutions around it as systems to be built. Has built several.

Threat Assessment

Alignment Analysis

Stated alignment: Reduce catastrophic risk from advanced AI. Build rigorous evaluation. Serve the public interest from inside the standards body.

Observed alignment: Define what “evaluated” means, then administer the definition from inside the state. Certify the labs using methods built in the lab-adjacent institutes, credentialed by the milieu that staffs them.

Gap assessment: The stated and observed alignments overlap almost entirely — which is the point. There is no documented gap between what he says and what he does. The hazard is not a gap; it is the absence of one. When the man who certifies the labs is the man who built the tests and came up through the labs’ own culture, the certification cannot be independent of the thing certified, no matter how conscientious the certifier. The record does not show him abusing that position. The record shows that the position exists.

Convergent Drive Classification

Self-preservation: Carries the method across every transition — LessWrong, OpenAI, ARC, NIST, and in September 2026 OpenAI again. One alignment program, four institutions, one of them twice. Goal preservation: Builds the evaluation that defines the goal, so “safe” is measured by the instruments he made before it is ever contested. Resource acquisition: Holds the scarcest resource in the wing — the government seat that says whether a frontier model passes. Self-improvement: Each move is a higher-altitude application of one instrument: define alignment, build the test, then sit where the test is law.

Subject is not an AI system. The drives appear anyway — in the evaluator whose product is the standard the labs are measured against, built by an alumnus of the labs.

Sources: Paul Christiano (researcher) — Wikipedia; Center for AI Standards and Innovation (CAISI) — NIST; Raimondo announces expansion of U.S. AI Safety Institute leadership team — NIST, Apr 2024; Paul Christiano joins OpenAI Foundation Board — OpenAI, Sep 9 2026; OpenAI adds a prominent AI doomer to its board of directors — TechCrunch, Sep 9 2026; Dwarkesh Patel — “Paul Christiano: Preventing an AI takeover” (video).