Daniel Kokotajlo: Why He Left OpenAI, and His Record

Daniel Kokotajlo is a former OpenAI governance researcher who co-founded and leads the AI Futures Project, a Berkeley nonprofit. He co-wrote AI 2027, a scenario about how fast AI could advance. Two questions follow him: why he left OpenAI, and whether his forecasts deserve trust.

Why Daniel Kokotajlo Walked Away From OpenAI

He quit OpenAI in April 2024, according to Euronews. He posted publicly that he left due to "losing confidence that [the company] would behave responsibly around the time of AGI." At his exit, he refused to sign OpenAI's non-disparagement clause, which could have cost him about $2 million in vested equity. He confirmed in May 2024 that he still had the equity.

He joined OpenAI's governance team in 2022. Wikipedia gives only the years, so this page uses 2022 until April 2024, per Euronews, and no more precise dates. The AI 2027 site describes his OpenAI work as "scenario planning." His training is in philosophy, not AI engineering.

Euronews, reporting on May 20, 2024, describes the clause as requiring a non-disclosure and non-disparagement agreement to keep vested equity. It says Kokotajlo "implied publicly on a blog that he gave up what would have been a large sum of money to leave without signing anything." Euronews doesn't give a dollar figure for him and doesn't say his equity was revoked.

Sam Altman said OpenAI had never "clawed back anyone's vested equity," and he said of the provision: "I did not know this was happening and I should have." Those comments were about the policy, not about Kokotajlo personally.

He declined to sign and accepted the risk of losing roughly $2 million in vested equity. OpenAI then changed the policy. As of May 2024 he still had the vested equity, so on these sources he did not permanently lose the money.

He was one of the organisers of a group of OpenAI employees who said the company had a secretive and reckless culture. They said it was taking grave risks in the rush toward AGI.

In June 2024 he signed a letter, with other current and former AI-company employees, calling for a "right to warn" about AI risks without fear of reprisal. The letter also protected anonymity.

What He Runs: The AI Futures Project

Kokotajlo is Executive Director of the AI Futures Project, according to the organisation's own about page. The project is a Berkeley, California nonprofit that researches the future impact of AI. He co-founded it and leads it. It is a 501(c)(3) with EIN 99-4320292, and it was formerly named Artificial Intelligence Forecasting Inc.

The about page doesn't publish the founding date. It says the project's funding "comes primarily from the Survival and Flourishing Fund and a major individual donor, along with other smaller charitable donations and grants." The donor isn't named, and the project doesn't publish a full donor list. The sources can't show whether AI-industry money reaches it.

The project's IRS Form 990 for fiscal year 2024 covers the year ending December 2024. It was filed November 13, 2025, and reports total revenue of $760,281 and total expenses of $226,188.

On July 9, 2026, the project published AI 2040: Plan A. It recommends that superintelligence development be delayed until 2040. The follow-up is described as more optimistic.

Our AI Futures Project explainer covers the group's work in more detail.

What AI 2027 Is, and What It Does Not Claim

AI 2027 is a scenario, not a single prediction. Its authors call it "the first major release from the AI Futures Project." It came out in April 2025, and both the New York Times article about it and ai-2027.com are dated April 3, 2025. Kokotajlo's method is scenario forecasting, which means writing detailed narrative futures instead of issuing point predictions.

The scenario branches. It has two alternative endings, one with continued competition and one with a slowdown. The authors invite pushback: "We encourage you to debate and counter this scenario."

It does not claim AGI arrives in 2027. A November 2025 clarification in the foreword says: "To prevent misunderstandings: we don't know exactly when AGI will be built." It adds that "2027 was our modal (most likely) year at the time of publication, our medians were somewhat longer."

The authors also disagree among themselves. Their about page says: "We disagree somewhat amongst ourselves about AI timelines; our median AGI arrival date is somewhat longer than what this scenario depicts." The stated aim is "to spark a broad conversation about where we're headed and how to steer toward positive futures."

If you want the broader timeline debate, our guide on how close we are to AGI covers it. For the wider group of people who warn about advanced AI, see our AI doomers explainer.

How His Earlier Forecast Held Up

The checkable part of his record is "What 2026 Looks Like," which he published on the AI Alignment Forum on August 6, 2021. He framed it as a future history written for the Vignettes Workshop, "as realistic (to me) as I can currently manage." He admitted he had to "make a bunch of stuff up," and he wrote: "I fully expect the actual world to diverge quickly from the trajectory laid out here."

Calls That Look Right in Hindsight

• By 2026, hundreds of millions of people would talk to chatbots regularly, and many would treat them as friends.
• Hype would peak in 2023, and the AI-risk community would shorten its timelines.
• From 2024, states including Russia would use language models to scale online propaganda and AI-driven astroturfing.
• AI would help design chips, and a new chip shortage would develop by 2026.

Calls That Missed or Were Walked Back

He predicted a 2026 "chatbot class consciousness," in which chatbots would talk virally about their own feelings. In a 2023 comment, he said he expected the big labs to restrict chatbot discussion of sentience and philosophy instead. He also predicted AI would reach human-expert level at Diplomacy in 2025.

The post also estimated the largest 2024 training run at about 5×10^25 FLOPs. The sources this page uses don't check that estimate against the real run. It also said his original draft's trends pointed toward AI takeover around 2027. He never finished that continuation, and that is where the "2027" framing traces back to.

The project's own about page says the 2021 forecast "held up well." That is the organisation's claim about its founder. No independent item-by-item audit of that forecast appears in the sources this page uses.

The record has one more dated revision. On December 31, 2025, the project's updated forecasts moved full coding automation to the early 2030s. That is a public move away from the original timeline.

Is Daniel Kokotajlo Credible?

His record has checkable hits and misses, and the timeline models behind AI 2027 have drawn detailed technical criticism.

The Case for Him

• He took a BA in philosophy at the University of Notre Dame in 2014, and an MA at the University of North Carolina at Chapel Hill in 2017, under John Thomas Roberts.
• He pursued doctoral studies in philosophy at UNC and held the 2019 to 2020 Maynard Adams Fellowship for the Public Humanities. The sources don't say whether he finished that PhD, so this page doesn't call him a doctor.
• Kevin Roose wrote in 2025 that of his 2021 forecast, "A number of his predictions proved prescient."
• TIME included him in its 2024 TIME100 AI list of the 100 most influential people in AI. The article is dated September 5, 2024.
• His AI 2027 co-author Eli Lifland "ranks first on the RAND Forecasting Initiative all-time leaderboard." That is a record for the team, not for Kokotajlo himself.

The Case Against Him

• Ali Farhadi, chief executive of the Allen Institute for AI, said the AI 2027 forecast was insufficiently grounded in scientific evidence.
• A detailed critique by the pseudonymous author titotal, published June 19, 2025, argues that the superexponential curve the authors give roughly 40 percent probability diverges. In that curve, "each successive doubling takes 10% less time."
• The critique says the curve "will still claim that superhuman coding will arrive before the end of 2029" regardless of starting values.
• It says a widely shared "AI 2027 prediction" graph omitted early GPT datapoints and misstated the decay as 15 percent, rather than 10 percent.
• It says two three-parameter fits to the METR data fit about equally well but diverge. One implies an infinite time horizon by 2030, and the other a horizon of a few months.

The authors engaged with the critique. Lifland, speaking for himself and Kokotajlo, agreed the tweeted graph "is not closely representative" of the model. He removed one argument, agreed the RE-Bench analysis was inconsistent, acknowledged communication problems, and awarded titotal a $500 bounty.

Lifland still defended the model as "the least bad" available. He said the updated model "still give[s] ~25-40% to superhuman coders by the end of 2027."

On money, his only disclosed financial interest in an AI company is the OpenAI vested equity that reporting says he retained as of May 2024. Reporting values that at about $2 million. The project's funding sources are named above, and its full donor list is not published.

Four points stay open. The $2 million figure is an estimate, not a confirmed valuation. The individual donor is unnamed. The sources don't show whether he finished his PhD. The organisation's praise for his 2021 forecast isn't independent. The timeline argument itself is unresolved, and the project has already revised its timeline once.

To judge the timeline argument yourself, read the authors' own framing on the AI 2027 about page, then the critique of its timeline models. For other voices in the AI debate, see our guide to who is fighting AI.

Sources

• Daniel Kokotajlo, Wikipedia
• What 2026 Looks Like, AI Alignment Forum (August 6, 2021)
• AI 2027 about page
• AI Futures Project about page
• AI Futures Project Form 990, ProPublica Nonprofit Explorer
• Euronews on OpenAI exit contracts (May 20, 2024)
• A deep critique of AI 2027's bad timeline models, LessWrong