AI Futures Project: Who Runs It and What AI 2027 Says

The AI Futures Project is a small nonprofit in Berkeley, California, that forecasts how artificial intelligence (AI) could develop. Its AI 2027 scenario is built on timelines models that project when AI systems become superhuman coders. The group's AI 2040 work includes a recommendation to policymakers in the United States (US).

What the AI Futures Project Is and What AI 2027 Says

The AI Futures Project is a nonprofit research organization co-founded and led by Daniel Kokotajlo, according to Wikipedia. It published AI 2027 on April 3, 2025, and the authors describe it as a scenario depicting their most likely case, not a central estimate.

The authors' own phrase is "something like our mode." A November 2025 clarification added to the foreword says the authors don't know when artificial general intelligence (AGI) will be built. It adds that 2027 was their modal year when they published, and their medians were somewhat longer.

The scenario has two alternative endings, one of continued competition and one of a slowdown. The about page and the Internal Revenue Service (IRS) record both describe the group as a 501(c)(3) nonprofit. It is registered in California and based in Berkeley, and because it holds that status, donations to it are tax-deductible.

Its later scenario, AI 2040: Plan A, was published July 9, 2026. The about page describes it as a scenario illustrating the group's recommendation to US policymakers, and it recommends delaying superintelligence development until 2040. For other groups working on limits to AI, see our guide to who is fighting AI.

Who Works There

The organization's about page lists eight named people as of October 2026. The team is small. Each person has a title and earlier affiliations listed on the page.

• Daniel Kokotajlo, Executive Director. Previously a governance researcher at OpenAI.
• Thomas Larsen, Research Lead. Previously founded the Center for AI Policy and worked on AI safety research at the Machine Intelligence Research Institute.
• Eli Lifland, Research Lead. Previously worked on Elicit and co-created TextAttack. Co-founded and advises Sage.
• Romeo Dean, Researcher. Previously an AI Policy Fellow at the Institute for AI Policy and Strategy.
• Lauren Mangla, Managing Director. Previously at Constellation and Executive Director of SPAR.
• Miles Kodama, Policy Researcher. Previously a summer research fellow at the Centre for the Governance of AI and a contractor at METR.
• Nicole Sanna, Chief of Staff. Previously built operations systems at the Atlas Fellowship, EA Funds and Sentinel Bio.
• Brendan Halstead, Research Contractor.

Kokotajlo works on scenario planning. Eli Lifland works on forecasting AI capabilities, and the page says he "ranks first on the RAND Forecasting Initiative all-time leaderboard." Thomas Larsen "previously founded the Center for AI Policy," and Romeo Dean "specializes in forecasting AI chip production and usage." Kokotajlo, Lifland, Larsen and Dean wrote the scenario and its endings.

The two sites disagree on one title. The organization's own about page lists Lauren Mangla as Managing Director. The AI 2027 about page lists her as COO, handling comms, hiring and tabletop exercises.

How AI 2027 Was Built

The AI 2027 about page describes three methods.

• Tabletop exercises. The team has run over 30 iterations of a tabletop exercise (TTX) to simulate AGI development. The scenario draws on more than a dozen of these exercises with hundreds of people. The page notes that "every TTX is different."
• Quantitative models. The site publishes a simplified timelines forecast, which the page calls a simplified version of the full AI 2027 timelines model. It also publishes a time-horizon trend graph and separate forecasts on compute, timelines, takeoff, AI goals and security.
• Outside forecasting. FutureSearch provided independent forecasting for the timelines, takeoff and revenue research.

Hundreds of people reviewed the text, and the page individually acknowledges roughly 75 to 80 reviewers. The about page says the models and graphs exist. It doesn't explain how the trend graph is built.

The site keeps a changelog of corrections. The AI 2027 about page records an error about the range of the authors' medians. The authors attribute it to "a mistaken understanding" of one co-author's view.

The group's track record claim is self-assessed. Its about page says Kokotajlo's earlier "What 2026 Looks Like" "held up well." No independent item-by-item scoring of that claim was found. For the wider debate over AGI timing, see our explainer on how close we are to AGI.

Who Funds It

The about page says funding "comes primarily from the Survival and Flourishing Fund and a major individual donor, along with other smaller charitable donations and grants." The organization doesn't publish a donor list.

The major individual donor isn't named. So whether any money from the AI industry is involved can't be established from its published sources.

The most recent public tax filing covers fiscal year (FY) 2024, which ended in December 2024. The Form 990 was filed November 13, 2025.

That return reports total revenue of $760,281 and total expenses of $226,188 for FY2024. Subtracting expenses from revenue leaves $534,093 more coming in than going out (760,281 minus 226,188). The sources don't explain that gap, so it is reported here as the filing shows it.

The exemption was issued in October 2024, so the FY2024 return covers a period in which the group had only just received it. The gap should be read as what the return reports. The IRS ruling year is also 2024.

The legal name on IRS records is Ai Futures Project Inc. The group previously operated as Artificial Intelligence Forecasting Inc. Its employer identification number (EIN) is 99-4320292. The ProPublica nonprofit record lists its filings, and that is the place to check them.

ProPublica's record doesn't show a Form 990 later than FY2024 as of October 21, 2026. The $760,281 figure is therefore the most recent public number, and it may be out of date.

The Substantive Critiques of the Scenario

The criticism targets the timelines models. Ali Farhadi, chief executive of the Allen Institute for AI, said the AI 2027 forecast was insufficiently grounded in scientific evidence. The most detailed critique is "A deep critique of AI 2027's bad timeline models", dated 19 June 2025 and posted on LessWrong.

The critic writes as titotal, a pseudonym. The source doesn't establish that person's identity or credentials, so the critique should be weighed on its arguments. For the broader risk debate, see our explainer on AI doomers.

The central objection concerns the superexponential curve. The authors assign roughly 40% probability to a curve in which "each successive doubling takes 10% less time." The critique argues that this formula diverges, so the curve still claims superhuman coding arrives before the end of 2029, whatever the starting values.

The critique lists further objections.

• Thin justification. The critique says the justifications for superexponentiality were thin. One of them, the gap between public and internal releases, pointed the other way.
• Backcast. A backcast implied current AI R&D is about 66% faster than in 2022. The appendix's own estimate is 3 to 30%.
• The public graph. The widely shared "AI 2027 prediction" graph left out early GPT datapoints. It stated the decay as 15% instead of 10%, and it didn't match the model's actual output.
• Overfitting. Two three-parameter fits to the METR data fit about equally well but diverge sharply. One implies an infinite time horizon by 2030, and the other a horizon of a few months.
• Benchmark ceiling. The upper bound of the RE-Bench logistic fit was guessed at 2.0. The critic argues that fitting the data without that assumption suggests saturation has already occurred, while a bound of 1.4 would make saturation take infinite time. The saturation date is set by the forecaster in the code, and the logistic fit is never used.

The authors engaged with the critique. Eli Lifland, writing for himself and Kokotajlo, agreed the tweeted graph "is not closely representative" of the model. He dropped the public-versus-internal argument, agreed the RE-Bench analysis was inconsistent, and acknowledged communication problems. The project also paid the critic a $500 bounty.

The authors kept several positions. They maintained that superexponentiality is plausible, that the RE-Bench ceiling is a reasonable estimate, and that intuitive adjustments on top of quantitative models are appropriate. They called the model "the least bad" available. They also conceded their median "could move significantly" after the critique's changes.

Even after the update, the authors still give roughly 25 to 40% to superhuman coders by the end of 2027. That figure is a probability.

The critic's own verdict on the May 2025 update is summarised below. It added uncertainty to the key parameter and made superexponentiality more complicated. It left most objections unresolved, and the expanded model was not yet documented enough to evaluate.

In the project's favor, it has moved its own numbers in public. Its December 31, 2025 update placed full coding automation in the early 2030s. AI 2040: Plan A followed on July 9, 2026.

Read the 2027 date as the authors' modal year, not their median. The critiques target the timelines models, and the authors have conceded their median could move. The next step is to read the critique and the response, both linked in the sources below.

Sources

• The AI Futures Project about page
• ProPublica nonprofit record for Ai Futures Project Inc.
• AI 2027 about page
• Wikipedia entry on Daniel Kokotajlo
• A deep critique of AI 2027's bad timeline models, on LessWrong