DANIEL KOKOTAJLO

Behavioral Archetype

THE FORECASTER WHO WALKED — Subject left a frontier lab’s governance team in 2024, declined an exit agreement that would have cost him his vested equity, signed a letter asking labs to stop gagging risk-related criticism, and built a small nonprofit whose product is a dated, falsifiable account of where the labs are heading. The authority is of an unusual kind. He runs no evaluator, sells no model access and writes no rule. He supplies the scenario that other people argue from, and he now supplies it to a Senate subcommittee, to a mission seat on an evaluator’s board and to an advisory line at a forecasting institute, all at once. The structural point is that the same person sits at three nodes of the evaluator map whose connecting tissue is a single philanthropic funder: the nonprofit he leads is funded primarily by Survival and Flourishing Fund, whose tables also fund the evaluators he advises. This profile scores the position, not the man.

Essence Indicators

• Executive Director, AI Futures Project. The project’s own About page says: “Daniel oversees our work. He previously worked as a governance researcher at OpenAI on scenario planning.” It describes the organization as “a 501(c)(3) nonprofit research organization (EIN 99-4320292)” (AI Futures Project, About). His written Senate testimony: “I’m the executive director of the AI Futures Project, a small nonpartisan nonprofit focused on forecasting the future of AI and recommending policies to address AI risks” (HSGAC testimony). The core team the home page lists is “Daniel Kokotajlo, Eli Lifland, Thomas Larsen, Romeo Dean, Lauren Mangla, Miles Kodama, and Nicole Sanna” (AI Futures Project).

• Former OpenAI governance researcher, 2022 to 2024. His own testimony: “I’ve been following the AI industry since before the founding of OpenAI and even worked there for two years,” and in a footnote, “While at OpenAI, I worked on a variety of projects including scenario forecasting, dangerous capabilities evaluations, and even a half-year stint on a capabilities team” (HSGAC testimony). His LessWrong profile gives the path: “Was a philosophy PhD student, left to work at AI Impacts, then Center on Long-Term Risk, then OpenAI” (LessWrong profile). TIME reports: “In 2022, Daniel Kokotajlo joined OpenAI as a governance researcher” (TIME100 AI 2024). The employer’s own profile is OpenAI.

• The resignation and the equity. In testimony: “I resigned in 2024 partly due to losing confidence that the company would behave responsibly, and partly so that I could speak more freely about the industry and where it is headed,” and “I also refused to sign OpenAI’s secret nondisparagement agreement, even though this would have cost me >80% of my family’s net worth. (In response to public uproar they changed the paperwork and I was able to keep the equity).” A footnote adds: “At the time, my vested equity was worth almost $2M” and “I’ve since sold almost all of it” (HSGAC testimony). In his own comment on a LessWrong thread he wrote at the time: “I did sign something when I joined the company, so I’m still not completely free to speak (still under confidentiality obligations). But I didn’t take on any additional obligations when I left,” and “Basically I wanted to retain my ability to criticize the company in the future” (LessWrong, comment thread). TIME reports the figure as “approximately $2 million in equity” (TIME100 AI 2024).

• Signatory, “A Right to Warn about Advanced Artificial Intelligence” (4 June 2024). The letter lists “Daniel Kokotajlo, formerly OpenAI” among its signers, beside “William Saunders, formerly OpenAI.” Its first commitment asks that a company “will not enter into or enforce any agreement that prohibits “disparagement” or criticism of the company for risk-related concerns, nor retaliate for risk-related criticism by hindering any vested economic benefit” (righttowarn.ai).

• First mission director, Apollo Research PBC. Apollo’s conversion post states: “These mission seats will be held by “mission directors” who are independent to Apollo Research PBC and our funders, and who have a special mandate that enables them to prioritize mission over other considerations. Our first mission director is Daniel Kokotajlo” (Apollo, 20 January 2026). The organization is profiled at Apollo Research, and its chief executive at Marius Hobbhahn.

• Research Advisor, Epoch AI. Epoch’s team page lists “Daniel Kokotajlo” with the title “Research Advisor,” in the advisor block beside “Miles Brundage,” “Strategy Advisor” (Epoch AI, team). Epoch’s own profile is Epoch AI.

• Author of the AI 2027 scenario. The AI 2027 site lists “Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean” and says “The AI 2027 scenario is the first major release from the AI Futures Project” and “We created this website in collaboration with Lightcone Infrastructure.” (AI 2027; AI 2027, About). His testimony names the earlier essay: “My forecasts have included “What 2026 Looks Like” which I wrote in 2021, and “AI 2027” which was published in April 2025,” and the later plan, “AI 2040: Plan A” (HSGAC testimony).

• The hearing. The committee lists Panel I of “Rogue AI: Securing the Homeland Against AI Agent Attacks,” 30 September 2026, with “Daniel Kokotajlo, Executive Director, AI Futures Project,” beside Chris Painter of METR and Marius Hobbhahn of Apollo, Paul Ohm and Kurt Gaudette (HSGAC hearing page). Tech Policy Press reports that he “described a future in which companies could put “AIs in charge of making the AIs that make the AIs,” reducing direct human supervision” (Tech Policy Press, 1 October 2026).

• TIME100 AI 2024. TIME’s entry quotes him: “A sane civilization would not be proceeding with the creation of this incredibly powerful technology until we had some better idea of what we were doing and how we were going to keep it safe” (TIME100 AI 2024). The project’s own pages link the same profile.

Social Persona / Impression Management

Immediate impression: The researcher who paid for the right to talk. Plain-spoken, numerate, a little rueful. The testimony opens by naming the cost before the argument: “In doing so I walked away from millions of dollars by abandoning an extremely well-paying job” (HSGAC testimony).

Energy: Forecaster’s register: trend extrapolation, dated predictions, then an explicit caution against his own method. “Forecasting is not an exact science by any means,” he writes, and, about the labs’ own timelines, “Company forecasts are subject to various biases and shouldn’t be trusted on their own” (HSGAC testimony). On a podcast, asked about the plan to hand alignment research to AI agents, he answers: “If you’re looking for someone to defend this plan, it’s not me. This is a terrible plan” (Palisade Research podcast, 25 August 2026).

Impression management strategy: The credible leaver. The exit story does the work a credential would: a refused agreement, a stated dollar figure, and a stated reason. He is also exact about what that story does not license. His own comment says he remained “under confidentiality obligations,” and his testimony states that the scenarios he made while at OpenAI “are not public.” He pre-empts the objection that his policy ask is a capture play: “I do worry that the leading AI companies will find a way to defuse that energy by making minor concessions that don’t really change the basic picture, or worse serve as a safety-washing function rather than real safety” (HSGAC testimony).

Forensic Archetype Comparison

Psychometric Assessment

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

Big Five (OCEAN):

Dark Triad (held low and evidence-bound; the score measures structural position, not character):

MBTI: INTP-adjacent: theory-first, builds the model and then tests it against the record. The exhibits: a 2021 essay that predicted the years 2022 to 2026 in advance, a stated method (“The main tool in the forecaster’s toolbox is trend extrapolation”) and a philosophy background (HSGAC testimony; LessWrong profile). Rationale is the public record, not a typing session.

Threat Assessment

Score move: none. Calibrated against peers with comparable seats. Jeffrey Ladish (7/6/7, hazard 58, Executive Director of Palisade Research) shares the insulation value for the same reason, a philanthropic funder base concentrated in one source, and sits above on leverage because his bio records advice to the White House, the Department of Defense and congressional offices where the subject’s record shows one Senate panel and an Apollo board seat. Marius Hobbhahn (7/7/7, hazard 60, CEO of Apollo) sits above on leverage because his organization is named in a lab’s system card and the subject’s is not an evaluator. Helen Toner (6/8/8, hazard 80, Interim Executive Director of CSET) shares the leverage value, a witness-and-author seat without authority over a deployed model, and sits above on insulation and durability: she led a university-hosted centre with a multi-year funded program and cast a vote that removed a chief executive, where the subject leads a three-year-old nonprofit funded by one fund and one donor. The other OpenAI-departure comparison is Jan Leike (7/8/8, hazard 76), who left the same lab in 2024 and held the alignment lead at another frontier lab; the subject holds no seat inside a lab, so atk 6, def 6 and hp 7 against his 7, 8 and 8.

Score basis. Atk 6 is leverage: his testimony recommends that the government mandate third-party access to incident logs, compute allocation and RSI-relevant information, and that it “intervene and require” the labs to slow, but a hearing does not enact either (HSGAC testimony); it is not a 7 because no document shows a lab, agency or rule adopting the ask. Def 6 is insulation: “Our funding comes primarily from the Survival and Flourishing Fund and a major individual donor” and he has “since sold almost all” of the OpenAI equity (AI Futures Project, About; HSGAC testimony); it is not higher because the funder is one fund and one donor. Hp 7 is durability: the right-to-warn letter and the scenarios are fixed public documents, and his testimony credits a change in a lab’s exit paperwork to public uproar (righttowarn.ai); it is not an 8 because none of them is a statute or program of record.

Alliance Map

The overlap is the finding. A Corporate seat as the executive director of a donor-funded nonprofit, a mission seat on an evaluator’s board and an advisory line at a forecasting institute, a Movement funding line that runs through one fund (SFF), and an Agency appearance before a Senate panel that also seated Apollo’s chief executive, all run at the same time, and they point at the same policy: outside, independent access to the labs, required by law. His own testimony names the risk that the market for that access fails, a “race to the bottom” in which third parties become “rubber-stampers that produce only safety-washing and red tape.”

Alignment Analysis

Stated alignment: Make the near-term trajectory of AI visible to the public and government, and slow the race to automated AI research until it can be controlled. In his words: “My message today is this: The leading AI companies are racing each other towards superintelligence,” and the remedy: “First, we need to dramatically improve transparency into the AI industry so that the public and government can adequately understand and respond to the firehose of new concerning developments” (HSGAC testimony).

Observed alignment: Consistent with the statement. He left a lab, kept the freedom to speak at a cost he puts at over 80 percent of family net worth, signed a letter asking for protected criticism, built a donor-funded forecasting nonprofit, and said the same thing to a Senate subcommittee that the project’s pages say. The testimony names both the benefit and the limit of an inside view: “the scenarios and forecasts I made while at OpenAI are not public” (HSGAC testimony).

Gap assessment: No documented gap between word and deed. The structural exposure is not his candour but his position on the map: the nonprofit he leads is funded primarily by one fund and one donor, the evaluator whose board he sits on appears in the same fund’s tables ($251,000 in SFF-2024, received through Rethink Priorities, and $315,000 in SFF-2026 to Apollo Research PBC), and so does the institute that lists him as an advisor (Epoch AI, $646,000 in SFF-2026) (SFF-2024; SFF-2026; see also Survival and Flourishing Fund). He is also exposed on the opposite flank: he states the regulatory-capture objection himself and answers it with a test, “we should be able to look back six months later and see a change in slope in the graphs of benchmark performance over time.” The record does not settle motive, and for the forecaster it never needs to.

Interest: (the faction’s stake in each position; a documented fact about the position, never a claim about motive.)

1. Position: mandate independent third-party transparency into the labs. “I recommend that the government mandate third-party transparency to at least the following information”: “Access all incident logs,” “Compute allocation,” and “Information relevant to measuring and forecasting recursive self-improvement” (HSGAC testimony).
   • What adoption would pay the subject’s seats. A required independent-auditor regime is a standing market for evaluators. The evaluator on whose board he holds the first mission seat says it “is moving from external monitor red-teaming to primarily focus on embedded evaluations” (Apollo). Item three of the same list, “Information relevant to measuring and forecasting recursive self-improvement,” is the subject matter of AI Futures Project’s own forecasts. No fee, contract or donation tied to the rule is documented for the project itself.
   • His own best case, steelmanned. “All of the incidents at OpenAI and at other AI companies should be investigated by credible, independent third parties, and investigated thoroughly instead of by three people over six days. And it should be required by law instead of something the companies can choose whether and how to do” (HSGAC testimony).
   • Who else gains from the same rule. Other outside evaluators (Marius Hobbhahn’s testimony asks Congress to require embedded evaluations; METR was the investigator he credits); legislators who want a regime to point at; labs that would obtain a binding standard to meet rather than a patchwork of voluntary ones.
2. Position: redirect compute away from automated AI research. “Recent estimates suggest that OpenAI and Anthropic both spend around 60-70% of their compute on AI R&D. If they decreased that to 10%, for example, this would significantly slow down AI capabilities progress towards RSI” (HSGAC testimony).
   • What adoption would pay the subject. No documented payment. The rule would make the project’s own question, how much compute goes to AI R&D, the number the government watches. His testimony says the compute allocation “should include how much compute is going towards AI R&D, towards alignment and control research, and towards serving AI products and services to external customers.”
   • His own best case, steelmanned. The freed compute “could then be spent on other things, including alignment and control research and, of course, providing AI products and services to customers, lowering prices for consumers,” and on China: “insofar as we are in a race with China, we are literally pulling them along behind us” (HSGAC testimony).
   • Who else gains from the same rule. Labs that are not at the frontier: his own testimony says “we should also see other companies begin to catch up to them somewhat in the market for coding agents.” Halt advocates and the funder base that backs them, which SFF’s tables also list.
3. Position: Plan A for the long run. The testimony points to “AI 2040: Plan A, a positive vision co-authored by me and others at AI Futures Project,” and on the podcast he describes its bedrock as “total research transparency” (HSGAC testimony; Palisade Research podcast).
   • What adoption would pay the subject. The project’s product, the scenario and the model of the future at aifuturesmodel.com, becomes the reference a policy answers to; the About page lists donations as a funding source (AI Futures Project, About). No contract is documented.
   • His own best case, steelmanned. “Plan A is complicated and fallible; it attempts to get the benefits of superintelligence while mitigating the risks. If you want a simpler plan, there’s always “Plan S:” Shut it all down” (HSGAC testimony).
   • Who else gains from the same rule. Any group that sells verification of training-run activity, and the governments that would gain a basis for treaty-style deals.

Nothing above is a verdict that the positions are wrong. The finding is the alignment of interests: the same person writes the scenario, leads the nonprofit that publishes it, holds a mission seat at an evaluator whose business a mandate would expand, and asks the Senate for the mandate, while saying publicly that the mandate could be captured.

Convergent Drive Classification

Self-preservation: Funded primarily by one fund and one donor, with smaller grants and donations; no lab pays. The OpenAI equity he once risked he has “since sold almost all of” (AI Futures Project, About; HSGAC testimony). Goal preservation: Holds the mission to the same sentence across venues: forecasting, transparency, slowing the race to automated research. The mission-director seat is a clause written into a charter so that it outlasts the board that wrote it (Apollo). Resource acquisition: Grants and donations to the project through SFF, with a matching pledge that SFF’s 2025 table prices at $500,000 with a 30 June 2026 deadline; a Senate platform; two advisory and governance seats (SFF-2025; HSGAC hearing page). Self-improvement: Each dated forecast is checked against events in public: the project’s own About page says his 2021 essay “held up well,” and the testimony tests its own claims against reported 2026 revenue and valuation figures (AI Futures Project, About; HSGAC testimony).

Subject is not an AI system. The drives appear anyway, in the forecaster whose scenario is the one the lab-facing evaluators and the legislators cite from the same panel.

Public footprint: X @DKokotajlo (linked from the AI Futures Project About page) · LessWrong author page · AI Futures Project · AI 2027 · written testimony at hsgac.senate.gov.

Sources: HSGAC — Rogue AI hearing page; Daniel Kokotajlo — written testimony, Senate HSGAC subcommittee, Sep 30, 2026; AI Futures Project; AI Futures Project — About; AI 2027; AI 2027 — About; Apollo Research is becoming a PBC; Apollo Research — homepage; Epoch AI — team; A Right to Warn about Advanced Artificial Intelligence; Daniel Kokotajlo — LessWrong profile; LessWrong — simeon_c’s Shortform (Kokotajlo comments); Palisade Research podcast — Daniel Kokotajlo on AI 2040 and Plan A; SFF-2024; SFF-2025; SFF-2026; TIME100 AI 2024 — Daniel Kokotajlo; Tech Policy Press — Senate Hearing Weighs Threats From Unrestrained AI Agents After OpenAI Hack, Oct 1, 2026.