Daniel Kokotajlo

Theory of Change

Kokotajlo's personal theory of change is to make the trajectory toward superintelligence visible and legible to policymakers and the public, creating political conditions for coordination over racing. He pursues this through four channels simultaneously:

Scenario forecasting -- making abstract risk concrete through detailed, date-specific narratives (AI 2027, the forthcoming AI 2030)
Tabletop exercises -- giving decision-makers visceral experience of AGI-era crises (35+ exercises with congressional staffers, lab researchers, journalists)
Whistleblowing and advocacy -- establishing accountability norms for AI labs (Right to Warn, amicus brief against OpenAI's for-profit conversion, SB 1047 support)
Policy development -- advocating specific mechanisms: hardware verification, transparency requirements, international coordination with mutual inspection

He is not trying to solve alignment directly. He is trying to buy time and change the political landscape so that alignment can be solved.

From the 80K Hours podcast (Jan 2026): "In the future, whoever controls all the AIs does not need humans. If you've only got one to five companies and they each have one to three of their smartest AIs in a million copies, then that means there's basically 10 minds that between those 10 minds get to decide almost everything."

What They Do

AI 2027 (April 2025): A 71-page scenario forecast reaching 1M+ readers. VP JD Vance and leaders of top AI companies reportedly read it. Written with Scott Alexander, Eli Lifland, Thomas Larsen, and Romeo Dean. Accompanied by five quantitative forecasts (compute, timelines, takeoff, AI goals, security). Informed by 30+ tabletop exercises. Reviewed by Carl Shulman, Helen Toner, Holden Karnofsky, Yoshua Bengio, and 60+ others.

Right to Warn (June 2024): Co-organized letter by former OpenAI employees calling for transparency. Endorsed by Bengio, Hinton, Russell. Directly influenced the AI Whistleblower Protection Act (bipartisan, introduced May 2025 by Grassley, Coons, Blackburn, Klobuchar). The AIWPA provides anti-retaliation protections, job restoration, double back pay, and prohibits contractual waivers of whistleblower rights.

OpenAI departure (April 2024): Resigned after "losing confidence in responsible AGI development." Refused non-disparagement clause, risking ~$2M equity (~85% family net worth). OpenAI reversed the policy; Altman called it "embarrassing."

Amicus brief (April 2025): Filed with 11 other former employees and Harvard professor Lawrence Lessig opposing OpenAI's for-profit conversion.

MATS mentorship (Summer 2026): Running an AI Futures Project stream at MATS, training next-generation researchers in scenario forecasting.

Forecasting track record: His 2021 "What 2026 Looks Like" anticipated chain-of-thought prompting, inference-time scaling, AI chip export controls, and $100M training runs before ChatGPT. His AGI median has shifted from 2027 (late 2022) to Dec 2030 (Jan 2026) in response to METR studies and model corrections.

Key People

Daniel Kokotajlo -- Executive Director, AI Futures Project. BA Philosophy (Notre Dame), MA Philosophy (UNC Chapel Hill). Career: AI Impacts (2019) -> CLR (~2020-21) -> OpenAI governance (2022-2024) -> AIFP (2025-). TIME 100 AI 2024 and 2025.

Eli Lifland -- Research Lead. #1 on RAND Forecasting Initiative all-time leaderboard. Co-founded Sage, worked on Elicit. His own AGI median is Jan 2035 -- significantly longer than Kokotajlo's.

Thomas Larsen -- Research Lead. Founded Center for AI Policy. Former MIRI researcher.

The core team is 5 people. Kokotajlo's position within the broader OpenAI safety exodus is significant: he was among the first to publicly resign, preceding departures by Jan Leike (Superalignment co-lead), Leopold Aschenbrenner, and eventually Mira Murati (CTO) and Ilya Sutskever. He reported that nearly half the ~30-person AGI safety staff had left by August 2024.

Money and Incentives

AI Futures Project: 501(c)(3) nonprofit, EIN 99-4320292 (previously Artificial Intelligence Forecasting Inc). Headquartered in Berkeley, CA. Needs $1.9-4.7M for 2026 operations.

Funding sources: Unknown. No grants found from Coefficient Giving, SFF, or any identified funder. Manifund page inaccessible. Accepts donations through every.org and DAFs. Website built by Lightcone Infrastructure; Scott Alexander volunteered writing; FutureSearch provided independent forecasting -- suggesting significant in-kind contributions from the rationalist/EA ecosystem.

Personal financial situation: Kokotajlo risked ~$2M OpenAI equity (retained after OpenAI reversed its NDA policy). He has skin in the game: offers monetary bets on predictions, pays bounties for errors found in his models ($500 to titotal, $500 to Peter Johnson).

Incentive analysis: Kokotajlo's incentives appear well-aligned with his stated mission. The equity sacrifice runs against financial self-interest. His main incentive risk is reputational: credibility depends on forecasting track record, which could bias toward defending past predictions. The recent 3-year timeline lengthening mitigates this concern. He does not receive funding from any AI lab.

What Others Say

Gary Marcus (external skeptic): AI 2027 is "a work of fiction, not science" and "a house of improbable longshots." More compellingly: it is "practically marketing materials" for AI companies and "puts wind in their sails." The China framing "feeds the worst fears of hawks, both in the US and China." Verdict: "Tall tales about the imminence of AGI aren't slowing down the AI race dynamic... they are speeding that very dynamic up."

Arvind Narayanan (Princeton, paradigm counter): "There is a long causal chain between AI capability increases and societal impact. Benefits and risks are realized when AI is deployed, not when it is developed." External bottlenecks (regulation, adoption, organizational change) cannot be overcome simply by improving AI's technical design, even by superintelligence. Acknowledges Kokotajlo's tech predictions were accurate but social impact predictions were "overall not directionally correct."

MIRI (insider critique): Agrees on direction but expects "more crazy, both in the sense of chaotic and in the sense of insane." The Slowdown ending is an "exercise in hope." The Race ending is "the only realistic path presented." Expects Agent-4 level models to "try to self-exfiltrate and set up dead-man's switches."

Vitalik Buterin: Accepts rapid capability gains as possible but argues AI 2027 "underrates defensive tech." In a world that cures cancer by 2029, defensive technologies also advance. Warns against the "one AI hegemon" strategy.

titotal (technical critic): Found specific bugs in timelines model, including a code error shifting the median by 9 months. AIFP acknowledged errors, paid $500 bounty, wrote 15K-word response.

Scott Alexander (collaborator): Notes team members have longer timelines than Daniel. "Skeptical" of very fast automation. Describes Slowdown ending as "exercise in hope."

Saffron Huang / Jessica Dai: AI 2027 "obscures human leverage and uses the authors' credentials to make doom seem inevitable."

What's Absent

AIFP funding sources: The single biggest factual gap. For an org seeking $1.9-4.7M that advocates transparency, the opacity is notable.
AIFP governance: No public board, no governance documents. Too new for 990s.
Current p(doom): The 70% figure (2024) is widely cited but likely outdated. If updated to 20-30%, this would be a significant shift.
OpenAI internal work: Kokotajlo's specific governance research outputs remain opaque.
Evaluation of 2025 predictions: The near-term predictions from AI 2027 should now be checkable, but no systematic public assessment was found.
Lab leader reactions: No specific public statements from Altman, Amodei, or Hassabis about AI 2027 were found.