# A forecast called AI 2027 predicts that superintelligent machines arrive in 2027 and the world ends shortly afterwards **No verdict.** Nothing here can be settled before the date arrives, and a page that pretended otherwise would be doing the same thing it is describing. What can be established now is narrower and still worth having. AI 2027 is a real document by named forecasters, published in April 2025, and it does contain a scenario in which humanity loses control. It also contains a second ending in which coordinated action slows things down and the disaster does not happen, and its authors state that the narratives are hypothetical planning tools rather than literal predictions. The single doom date that reached the public is a compression of something with a branch in it. Whether fast timelines are right is a live question among serious people, and this page does not pretend to answer it. Category: Science, Space & Technology · Era: 2020s · First circulated: April 2025, when the scenario was published, and spreading well beyond its original readership through 2025 and 2026 as the phrase AI 2027 detached from the document and became a date · Believed by: Hard to measure, and the search traffic is the best available proxy: people type 'is AI 2027 on track', 'is AI 2027 realistic', 'is AI 2027 coming true' and 'will the world end in 2027 AI'. The questions are asked in a register of genuine worry rather than of enthusiasm URL: https://theconspiratory.com/theory/ai-2027-scenario ## Summary In April 2025 a small nonprofit called the AI Futures Project published a scenario titled AI 2027. It reads like fiction and was built like a forecast: month by month from 2025, through AI systems that can do the work of AI researchers, to an intelligence explosion in late 2027. It got read by people who do not usually read forecasting documents, including reportedly the US vice president, and then the title came loose from the thing. What circulates now is a year. People search for whether AI 2027 is on track, whether it is realistic, whether the world ends in 2027, and a fair number of them have never opened the document. Two things in it complicate the version that travels. The scenario has two endings and one of them is fine, and the authors say plainly that they are not predicting, they are mapping a path that could plausibly happen. This file is about the distance between those two objects: a branching planning exercise, and a date. ## The claim That AI 2027 is a prediction, by credible insiders, that artificial superintelligence will arrive in 2027 and will probably destroy or irreversibly transform human civilisation, and that the world is currently on track for it. ## Origin and timeline - 2021-08: Daniel Kokotajlo, then working in AI forecasting, publishes an essay titled 'What 2026 Looks Like', a year-by-year sketch of how AI might develop. It is unusual for being specific enough to be graded later, which almost no forecast of this kind is. - 2024-04: Kokotajlo resigns from OpenAI's governance division, saying he had lost confidence that the company would behave responsibly around the arrival of very capable AI. He goes on to found the AI Futures Project, a US nonprofit funded by donations. - 2025-04-03: The AI Futures Project publishes AI 2027, written by Kokotajlo with Eli Lifland, Thomas Larsen and Romeo Dean, and edited by the blogger Scott Alexander. It describes developments month by month from 2025: AI agents in 2026, effectively automated coding by early 2027, and an intelligence explosion in late 2027 as systems begin improving themselves. - 2025-04: The scenario ends twice. In one branch international competition drives development past the point of human control, with catastrophic results. In the other, coordinated action slows the race down and the worst is averted. The authors describe both as hypothetical, and the project as a planning tool rather than a forecast of what will happen. - 2025-04-09: The Machine Intelligence Research Institute publishes a response that is sympathetic to the conclusion and sceptical of the method. Its author agrees that catastrophe is plausible on roughly this timescale, and writes that as far as he can tell nobody associated with AI 2027 actually expects events to move as fast as depicted, the scenario being a story about how things could plausibly go fast rather than a claim that they will. - 2025-04-23: Scott Alexander publishes a round-up of media coverage, reactions and criticism on the project's own blog, which is itself a sign of how far the document had travelled in three weeks. - 2025-12: The authors publicly revise. They write that developments since publication suggest a slower path to fully autonomous AI research than the scenario assumed. - 2026-04-14: Kokotajlo publishes an assessment of his own 2021 essay in Asterisk magazine, grading how the 2026 he sketched five years earlier compares with the one that arrived. Whatever the score, the act is the point: forecasts of this kind are almost never marked. - 2026-04: The authors revise again, in the other direction, writing that further developments in AI programming tools suggest a pace closer to the original scenario. They state an intention to keep integrating new evidence as it arrives. - 2026-07-09: The AI Futures Project publishes AI 2040: Plan A, a policy follow-up arguing for slowing the research timeline, forcing more transparency from AI labs, and spreading power across more companies and countries. The new number in the title is a target for policy, not a rescheduled apocalypse. ## The evidence, claim by claim - Claim: AI 2027 predicts the end of the world in 2027. It is in the title. Evidence: The document has two endings and only one of them is bad. In the other, international coordination slows the race and the catastrophe does not occur. A title is not a conclusion, and this one names the year the scenario begins to turn rather than a date of death. The authors also state that both narratives are hypothetical and that the project is a planning tool rather than a literal forecast, which is an unusual thing to have to point out about a document people are frightened of. - Claim: The people who wrote it are insiders. One of them left OpenAI over safety. That is why it should be believed. Evidence: Kokotajlo's resignation is real and his reasons were stated publicly, and it is reasonable to weigh the testimony of someone who walked away from a job over what he saw. But provenance is not prediction. The same author published a gradeable forecast in 2021 and went back in 2026 to mark it himself, which is the behaviour of a forecaster rather than a seer, and it cuts both ways: it invites you to check the record rather than to trust the credential. - Claim: The date is the consensus. Serious people expect superintelligence in 2027. Evidence: Not quite, and the distinction is the most useful thing on this page. Aggregated prediction markets at the time peaked in or near the end of 2027, and their midpoint sat in 2031. The most likely single year and the expected year are four years apart, because the distribution has a long tail to the right. 2027 is the mode. It is not the forecast. A number that is memorable because it is the tallest bar on a chart makes a poor deadline. - Claim: It is already happening. The 2026 milestones came true. Evidence: The authors themselves say the record is mixed, and they have moved in both directions: slower than the scenario in December 2025, closer to it in April 2026. Critics assessing measured progress have put the real pace at roughly two thirds of the scenario's. What nobody can do yet is grade the part that matters, because the scenario's decisive claims sit in late 2027 and the year has not happened. - Claim: It was built step by step from plausible increments, so the conclusion must be plausible too. Evidence: This is the sharpest criticism and it is not a dismissal. The method was to imagine the most plausible next step, repeatedly. MIRI's response names the failure mode that invites: the planning fallacy, familiar to anyone who has ever renovated a house or shipped software. Real intervals contain interruptions, and a chain of individually plausible steps can produce an implausible whole. The authors' answer is that critics should say which specific step is wrong, which is a fair demand and a hard one. - Claim: It has no date in it at all, so there is nothing to check. Evidence: It has a year and not a day, which matters more than it sounds. Every dated prophecy this site covers names a day, and a day is the one thing that answers itself: it arrives, and then you know. A year is softer. It can be read as underway for twelve months and as nearly over for the last of them, and it expires quietly rather than failing. That is not a criticism of the authors, who never offered a day. It is an observation about what survives when a document becomes a number. ## Why people believe it - The source is credible and the credibility is earned. A governance researcher resigned from a leading AI company over what he believed it would do, then founded an organisation to forecast the consequences in public. That is not the profile of somebody selling an apocalypse. - The worry underneath is shared by people with nothing to sell. Critics who reject the method still put catastrophe within a decade, and the aggregate of prediction markets has its peak in 2027. Someone who reads only that the tallest bar is 2027 has not misread the chart, only stopped early. - The document is specific, and specificity reads as knowledge. Month-by-month detail, named capabilities and a described mechanism feel like the output of a process, where a vague warning feels like an opinion. Detail is a property of the writing rather than of the future. - The subject changes visibly, every few months. Anyone using these tools can see capability moving, and extrapolating a curve you can personally feel is far more persuasive than extrapolating one you have to be told about. - A number is portable and a caveat is not. 'AI 2027' fits in a headline, a search box and a conversation. 'A two-branch scenario intended as a planning exercise, whose authors have revised it in both directions' does not, so the first travels and the second stays behind. ## Open questions - Will the late-2027 claims be gradeable when the time comes? An intelligence explosion is not the kind of event that announces itself on a Tuesday, and the scenario's decisive milestone is defined by capability rather than by anything a newspaper would report. It is possible for the year to pass without agreement on whether it happened. - How did the 2021 essay actually score? Kokotajlo marked 'What 2026 Looks Like' in April 2026, which makes it one of very few AI forecasts ever assessed against the world by its own author. It deserves reading before anyone decides how much weight the 2027 document carries. - Do the public revisions continue? Two came in four months and they pointed opposite ways. A forecast that keeps moving with the evidence behaves like a forecast; one that only ever slips forward behaves like something else. Which of those AI 2027 turns out to be is answerable, and not yet answered. - What happens to the phrase after the year ends? The document will still exist, its authors will presumably still be working, and 'AI 2027' will be a date in the past attached to an argument about the future. Whether it is retired, re-dated or quietly kept in circulation is the thing to watch, and it is the question this file will return to. ## Sources - AI 2027, AI Futures Project (2025): https://ai-2027.com/ - Our first project: AI 2027, AI Futures Project (2025): https://blog.aifutures.org/p/our-first-project-ai-2027 - AI 2027: Media, Reactions, Criticism, AI Futures Project (2025): https://blog.aifutures.org/p/ai-2027-media-reactions-criticism - Thoughts on AI 2027, Machine Intelligence Research Institute (2025): https://intelligence.org/2025/04/09/thoughts-on-ai-2027/ - Before he wrote AI 2027, he predicted the world in 2026. How did he do?, Asterisk (2026): https://asteriskmag.substack.com/p/before-he-wrote-ai-2027-he-predicted - AI Futures Project, Wikipedia: https://en.wikipedia.org/wiki/AI_Futures_Project Rated by The Conspiratory, a neutral, sourced encyclopedia of conspiracy theories. Full page: https://theconspiratory.com/theory/ai-2027-scenario