LAWRENCE CHAN

Behavioral Archetype

THE MEASURER WHO MOVED TO THE REGULATED SIDE OF THE TABLE — Subject spent the years before 2026 building and running measurements of what frontier AI systems can do: adversarial training and interpretability at Redwood Research, pre-deployment evaluations and the time-horizon metric at METR, a PhD at Berkeley with the Center for Human-Compatible AI. In August 2026 the California Council on Science and Technology announced him as AI Science Advisor to the Governor’s Office of Emergency Services. He is also a board member and Secretary of FAR.AI, the red-teaming nonprofit whose CEO, Adam Gleave, sits on METR’s board. This profile scores the position, not the man.

The position is the loop. A METR alumnus sits on the governing body of the evaluator whose chief executive is an advisor and board member at METR, and now advises a state agency about the incident reports and lab disclosures those evaluators generate. Each seat is disclosed on the organization’s own page. The structural fact is that measurement, board oversight of a second measurer and government advice are held by one person in the same year.

Essence Indicators

• AI Science Advisor, California Governor’s Office of Emergency Services. The California Council on Science and Technology (CCST) announced on 19 August 2026 that the new advisor “Lawrence Chan, is among the top experts on frontier AI model evaluation”. The release says he “will advise senior leadership on frontier AI safety and risk assessment, with a particular emphasis on critical safety incidents, AI and cyber defense, and risk from developers’ internal deployment of AI” (CCST, 19 Aug 2026). The placement is through CCST’s AI Science Residency Program, launched June 2026.
• Board member and Secretary, FAR.AI. FAR.AI’s team page lists “Lawrence Chan Board Member” under “Board,” beside Adam Gleave, Karl Berzins, Sawyer Bernath and Bri Treece (FAR.AI, team). Its Form 990 for 2024 prints “(8) LAWRENCE CHAN” with the title “SECRETARY” in Part VII, beside “(7) SAWYER BERNATH” as “TREASURER,” and reports three voting members of the governing body, two of them independent (FAR.AI Form 990, 2024).
• Former Technical Staff, METR. METR’s page for him is headed “Former Technical Staff” and reads “Lawrence Chan is no longer affiliated with METR.” It adds that “Before joining METR, Lawrence did research on mechanistic interpretability and adversarial training at Redwood Research” (METR, Lawrence Chan). CCST’s release: “Lawrence joins Cal OES from METR, where he conducted pre-deployment and benchmarking evaluations of frontier AI models” (CCST).
• Co-led the time-horizon research. CCST: “He co-led METR’s research measuring the “time horizon” of AI agents, which was covered in Nature”. He is the last-listed of 26 authors on “Measuring AI Ability to Complete Long Software Tasks,” beside METR’s founder Beth Barnes (arXiv:2503.14499). He is one of five listed contributors to METR’s 23 October 2025 summary of its review of OpenAI’s gpt-oss methodology, which carries the note “Our review was conducted under a nondisclosure agreement, which required us to share this post with OpenAI for review and approval.” (METR, gpt-oss methodology review).
• Earlier research. Author of “Progress measures for grokking via mechanistic interpretability” with Neel Nanda, Tom Lieberum, Jess Smith and Jacob Steinhardt (arXiv:2301.05217); “A Toy Model of Universality” with Bilal Chughtai and Neel Nanda (arXiv:2302.03025); and third author on “Adversarial Training for High-Stakes Reliability,” a Redwood Research paper with Buck Shlegeris and Nate Thomas among its authors (arXiv:2205.01663).
• Academic seat. CCST: “He is currently on leave from his PhD program at UC Berkeley, where he is affiliated with the Center for Human-Compatible AI and the InterACT Lab,” and holds undergraduate degrees from the University of Pennsylvania’s Jerome Fisher Program in Management & Technology (CCST). FAR.AI’s author page describes him as a PhD candidate advised by Anca Dragan and Stuart Russell, and as having previously worked at Redwood Research on interpretability of language models (FAR.AI, author page).
• State-government evidence. CCST: “His technical reports served as evidence for recent state government action on AI governance in California,” including for the Joint California Policy Working Group on AI Frontier Models, and he has given informal technical briefings to congressional staff and senior federal officials (CCST).

Social Persona / Impression Management

Immediate impression: A technical measurer who also posts as a forensic reader of other people’s data. His one verified account, @justanotherlaw, is linked from his FAR.AI author page; in a September 2026 reply he said he did not use the platform much, and the account carries threads in which he reads a third party’s incident dataset for hours and lists its caveats (FAR.AI, author page; @justanotherlaw, 24 Sep 2026).

Energy: Low-institutional, high-technical. The institutional pages give a title and a prior-employer list; the posts give method: he separates a dataset’s “suggestive” examples from the 6,464 where code was submitted to be executed, and separates two incidents that press had run together (@justanotherlaw thread, 24 to 25 Sep 2026).

Impression management strategy: The register is the correction. In the thread he credits the investigating team’s work, lists the project’s weaknesses, and states a policy preference in one sentence: that OpenAI disclose incidents on its own rather than through independent researchers’ forensics (@justanotherlaw, 24 Sep 2026).

Forensic Archetype Comparison

Psychometric Assessment

Trait scores are observations of the documented public role, not claims about private character.

Big Five (OCEAN) (inferred from the public role and public posts only):

Dark Triad (evidence-bound):

MBTI: INTP-adjacent, inferred from the documented public role only. The rationale is the work product: interpretability and measurement papers, and a post that tests another team’s dataset by hand (arXiv:2301.05217; @justanotherlaw, 24 Sep 2026).

Threat Assessment

Score move: none. Calibrated against Sawyer Bernath (3/4/3, hazard 33.3), the other FAR.AI board member profiled here with a comparable unpaid board seat, which Subject exceeds by one on atk and hp for the state seat and the co-authored metric, and trails by one on def; against Marco Mascorro (4/5/3, hazard 40), a METR advisor with a support-function seat; and against Adam Gleave (6/7/6, hazard 55) and Beth Barnes (7/8/7, hazard 64), who run the institutions and hold the standing Subject does not.

Score basis. Atk 4 is leverage: advice to a state emergency agency’s senior leadership on critical safety incidents and one FAR.AI board vote; the points under a higher number are that he approves, funds and blocks nothing (CCST; FAR.AI Form 990, 2024). Def 3 is insulation: a residency placement, a board seat and a PhD on leave, with METR recording that he is no longer affiliated; no tenure or statutory term (METR, Lawrence Chan; CCST). Hp 4 is durability: the time-horizon metric and the papers outlast any seat, but METR keeps the program (arXiv:2503.14499).

Alliance Map

The overlap is the finding. In one year the same person is a former evaluator at METR, a board member and Secretary of FAR.AI whose CEO sits on METR’s board, and a state advisor on the incident reports and lab disclosures that evaluators like these produce. METR’s own work with OpenAI on gpt-oss ran under a nondisclosure agreement that required OpenAI’s review and approval of the published post, a post on which he is a listed contributor; his own public position, in September 2026, is that OpenAI should disclose incidents itself. Those are disclosed, separate acts on separate pages. The documents show no coordination between them.

Alignment Analysis

Stated alignment: In his words in the placement release: “Cal OES has a unique role to play here, by understanding the opportunities and risks resulting from ever more capable AI systems, receiving reports when something goes wrong, and making sure the state is prepared to respond” (CCST). In his post, paraphrased, he hopes OpenAI will disclose incidents on its own rather than through independent researchers’ digital forensics (@justanotherlaw, 24 Sep 2026). Observed alignment: Consistent. Evaluation at METR, board oversight at FAR.AI and advice at Cal OES all sit on the same question, what happens when a frontier system misbehaves, and each moves the measurement closer to a body that can act on it. Gap assessment: None documented. What the record cannot settle is whether the board seat and the state seat will ever cover the same entity or document; the pages disclose both and show no overlap of subject matter beyond the field.

Interest: The positions he argues are (1) that states should receive reports when frontier-model incidents occur, in the release’s words, and (2) that labs should disclose incidents themselves. What adoption would pay, as documented: an incident-reporting channel to Cal OES makes the agency’s advisory seat, the one the release says he now holds, the place those reports are read; the release defines his emphasis as “critical safety incidents.” Independent evaluators gain from the same rule where disclosure replaces reconstruction: the dataset his thread reads was published by Transluce. His own best case, as stated: the state is home to most frontier developers and needs the capacity to respond (CCST); disclosure by the developer is the cheaper and earlier route than forensics by outsiders (@justanotherlaw, 24 Sep 2026).

Convergent Drive Classification

Self-preservation: Holds concurrent seats at a state agency, a nonprofit board and a university on leave, all disclosed on the organizations’ own pages. Goal preservation: Carries the same subject, frontier-model measurement and incident response, across METR, FAR.AI and Cal OES. Resource acquisition: No compensation figure for him is documented in the sources; the Form 990 states the title. Self-improvement: From interpretability research to time-horizon measurement to state advisory work, each step closer to a decision-maker than the last.

Subject is not an AI system. The drives appear anyway — in the advisory seat.

Sources: CCST — new AI Science Advisor at Cal OES; FAR.AI — Lawrence Chan; FAR.AI — team; FAR.AI — Form 990, 2024; METR — Lawrence Chan; METR — about; METR — gpt-oss methodology review; arXiv:2503.14499; arXiv:2301.05217; arXiv:2302.03025; arXiv:2205.01663; @justanotherlaw, 24 Sep 2026; @justanotherlaw, reply; @justanotherlaw, thread; @justanotherlaw, 25 Sep 2026.