The AI Futures Project is a nonprofit AI forecasting and scenario-planning research group founded by Daniel Kokotajlo after his April 2024 departure from OpenAI. The group is best known for *AI 2027* (April 2025), a month-by-month scenario forecast of AI development through a 2027 intelligence explosion that became one of the most-discussed AI-forecasting documents of 2025–26. Its subsequent publications include a quantitative timelines-and-takeoff model (the AI Futures Model), a self-assessment grading AI 2027's 2025 predictions, and AI 2040: Plan A (July 2026), a policy-recommendation scenario depicting a negotiated international slowdown of superintelligence development.
Provenance note: citations to "Grading AI 2027's 2025 Predictions" below are inline URL citations pending ingest.
Publications and work
AI 2027 (April 2025). The group's first major publication, released April 3, 2025 at ai-2027.com by authors Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean, projects AI development month-by-month from 2025 through a forecast 2027 intelligence explosion, branching into a "Race" ending (AI takeover) and a "Slowdown" ending (human control retained) (AI 2027).
AI Futures Model. The group maintains a quantitative model of AI timelines and takeoff speeds, published at aifuturesmodel.com, with a significant update released December 30, 2025 by Kokotajlo, Eli Lifland, Brendan Halstead, and Alex Kastner (Source: blog.aifutures.org). The group also publishes policy analysis on its blog, including "Early US policy priorities for AGI" (December 2025) by Nick Marsh (Source: blog.aifutures.org).
Grading AI 2027's 2025 Predictions (February 2026). On February 12, 2026, Eli Lifland and Kokotajlo published a self-assessment of AI 2027's 2025 predictions, finding that quantitative progress was tracking at roughly 65% of the scenario's predicted pace — with SWE-bench Verified progress "surprisingly slow," METR coding time horizons at 1.04× of a central AI-2027-speed trajectory, and OpenAI revenue slightly ahead of prediction — implying, if the pace held, a takeoff shifted to roughly mid-2028 through mid-2030 rather than 2027 (Grading AI 2027's 2025 Predictions (AI Futures Project, February 2026)). The post also disclosed an error in the trajectory graph published with AI 2027, against which the same METR result reads as 0.66× rather than 1.04×, and volunteered Grok's "MechaHitler" episode as a possible counterexample to the scenario's prediction that no post-2024 deployment incident would be as extreme as the Gemini and Bing Sydney cases.
AI 2040: Plan A (July 2026). On July 9, 2026, the group published AI 2040: Plan A at ai-2040.com, framed explicitly as a recommendation rather than a prediction: a scenario in which humanity delays superintelligence development to around 2040 through a US–China "verified slowdown" agreed in 2029, total research transparency for AI R&D ("everything but the model weights"), a pause at top-human-expert-level AI from 2035, and wider distribution of AI capability across dozens of companies and countries (AI 2040: Plan A (AI Futures Project, July 2026)). The scenario's own summary is that "humanity delays the development of superintelligence until 2040, makes all AI research public, allows dozens of companies globally to catch up to the frontier, and intentionally enters a regime of mutually assured compute destruction," with the transparency provision doing the enforcement work: total research transparency "allows the nations of the world to understand what's happening and enforce guardrails" (AI 2040: Plan A (AI Futures Project, July 2026)). The scenario contrasts Plan A with four alternative plans (B, C, D, and S) corresponding to other possible US responses, and is accompanied by supplements on transparency and verification mechanisms. The book-length scenario's six authors include Kokotajlo, Eli Lifland, AI 2027 co-author Thomas Larsen, and Redwood Research's Ryan Greenblatt (Source: axios.com; ai-2040.com); coverage noted its contrast with the group's earlier, more pessimistic scenario work (Source: washingtonpost.com). AI 2027 co-author Scott Alexander published a companion essay introducing the scenario the same day, and the group solicited further critique on July 10 (Source: astralcodexten.com; semafor.com). On July 10, Larsen published a follow-up research agenda calling for gamed-out scenarios of competing plans (an indefinite halt "Plan S," GPU arms control, a CERN for AI), better modeling of covert projects and verification, and economics work that takes AGI-driven growth seriously (Source: blog.aifutures.org). On July 11, Zvi Mowshowitz published an introduction and reaction roundup describing Plan A's core as a mutually verifiable U.S.–China deal to slow AI development — with joint control over existing and new chip supply and universal auditing of data centers — and calling it "the single most thorough and thoughtful plan" available for approaching superintelligence; Kokotajlo highlighted total research transparency as the element he weights most (Source: thezvi.substack.com). The same day, programmer George Hotz published a critique of the scenario (Source: ai-2040.com).
Methodology
The group's stated method is "scenario scrutiny": writing detailed, concrete, falsifiable scenarios as a vehicle for stress-testing forecasts and policy proposals, then grading them against reality as time passes (Grading AI 2027's 2025 Predictions (AI Futures Project, February 2026)). AI 2040: Plan A applies the same method to the group's own policy recommendations, arguing that most AI policy proposals "fall apart under scenario scrutiny" and that the discourse would improve if proposals were routinely subjected to it. The group's explanation for the practice's rarity is an incentive account — scrutinizing one's own proposals "might surface uncomfortable issues," while scrutinizing rivals' is "a lot of work for little rhetorical gain" — and it asks to be judged "against the existing state-of-the-art for plans to navigate the AI transition (if they can find any)" rather than against "some hazy but pleasant fantasy" (AI 2040: Plan A (AI Futures Project, July 2026)).
Positions
Across its publications the group argues that AI companies will probably succeed at building smarter-than-human systems within roughly one to ten years; that a full-speed race to superintelligence risks either loss of human control or an unprecedented concentration of power in whoever controls the leading systems; and that an international agreement with verification — rather than unilateral restraint or pure acceleration — is the preferable response (Source: ai-2040.com). Its near-term policy recommendations include limiting the gap between internal and external model deployment, requiring public model specifications, enforcing existing export controls, investing in verification R&D, and building government AI capacity. The export-control recommendation carries an explicit reservation: the group has "major reservations about introducing new export controls because they exacerbate the US/China race," but holds that existing ones should be enforced or else repealed, since smuggled chips — roughly a third of Chinese compute on Epoch's estimate — "make future agreements based on compute governance more difficult to enforce." Two further measures are capping the share of compute spent on AI R&D, around half of frontier-lab budgets in 2026, and ending AI-chip recycling, since decommissioned chips are "one of the most promising routes for covert projects to acquire chips later" (AI 2040: Plan A (AI Futures Project, July 2026)).
The threat model has two independent limbs. On control, the group holds that the industry "has convinced itself that controlling superintelligent AI can be figured out on the fly, and thus has no remotely adequate plan," and that no racer will have much of a lead or unilaterally slow down. On concentration, the argument does not depend on alignment failing: "even if the AI companies somehow align their AIs, the result will be an unprecedented concentration of power" in "a tiny group of people, or possibly just a single individual" (AI 2040: Plan A (AI Futures Project, July 2026)).
Relationships
- depends-on: Daniel Kokotajlo (founder)
- related: AI 2027 (the group's first major publication), AGI Timelines, Recursive Self-Improvement (RSI), AI Race Dynamics
- related: Broad Timelines (alternative timelines framework)