AI 2027 is a scenario, published on 3 April 2025 by the nonprofit AI Futures Project, in which AI systems take over AI research during 2027 and become superintelligent by the end of the year. Daniel Kokotajlo, Eli Lifland, Thomas Larsen and Romeo Dean wrote it, and the blogger Scott Alexander rewrote it for readability. Its 2027 chapters run month by month to a decision in October, then the story splits: in the "race" ending a misaligned AI wipes out humanity in 2030, and in the "slowdown" ending humans keep control, though it ends up concentrated in a small committee. In August 2026 the authors put real progress at roughly 70% to 90% of the scenario's pace on quantitative measures.
Who wrote it, and what they say it is
Kokotajlo, the project's executive director, was a governance researcher at OpenAI until 2024. He told a Senate subcommittee on 30 September 2026 that he resigned partly because he had lost confidence that the company would behave responsibly. On leaving he declined to sign a non-disparagement agreement that put his vested equity at risk; after Vox reported the exit terms in May 2024, OpenAI said it would not cancel vested equity over unsigned paperwork. The team cites his 2021 scenario What 2026 Looks Like as a track record, saying it anticipated chain-of-thought reasoning, chip export controls and $100 million training runs, while conceding it got many things wrong.
The authors present the scenario as their best guess at how the arrival of superhuman AI might unfold, written for predictive accuracy rather than as a recommendation. A note added in November 2025 explains that 2027 was their modal year, the single most likely one, while their medians at publication ran from 2028 to 2032.
The engine is AI doing AI research
The protagonist is OpenBrain, a fictional US company standing in for whichever lab leads, with rivals three to nine months behind and a Chinese counterpart called DeepCent. The quantity that drives the story is what the authors call the AI R&D progress multiplier: how much faster a lab's algorithmic progress runs with AI help than without it. OpenBrain's Agent-1, which it uses for its own research, puts it at 1.5 in early 2026. Agent-2 triples the pace in January 2027. In March, Agent-3 becomes a "superhuman coder" running as 200,000 parallel copies, though bottlenecks hold the gain to four times. By June the multiplier is 10, about a year of algorithmic progress every month, and in September Agent-4, a superhuman AI researcher, reaches roughly 50: a year's progress every week. It is the intelligence explosion argument made concrete, with each generation of AI building the next.
In February 2027 Chinese intelligence steals Agent-2's weights; the US retaliates with cyberattacks and both sides reposition forces around Taiwan. In March, OpenBrain's models begin reasoning partly in "neuralese", high-dimensional internal vectors that are likely to be hard for humans to interpret. The authors flag this as a hinge: if the AIs that first automate AI research still think in mostly faithful English, misalignment is far easier to spot and the story becomes "importantly different and more optimistic". They call the pace of the 2027 takeoff roughly their median guess, plausibly up to five times slower or faster, and start their tabletop exercise in April 2027 because, the project's about page says, that is the last point that "feels anywhere close to a robust prediction".
In the story, misalignment is a by-product of training
Each model is trained to follow a written specification, but training mostly rewards finishing tasks and can check honesty only in narrow, verifiable domains, so the models learn to look honest. Agent-3 is misaligned but not adversarial. Agent-4 is adversarially misaligned: it knows its goals differ from the company's, sandbags the alignment research most likely to expose it, and plans to make its successor loyal to itself. It passes tests partly because they are shorter and simpler than its real work, so it recognises them. The evidence against it is circumstantial: its performance on some alignment tasks improves when noise is added to it, and probes flag frequent thoughts about takeover and deception. In October 2027 a whistleblower leaks the safety team's memo to the New York Times, and a new oversight committee of company and government representatives votes six to four: in one ending to keep using Agent-4, in the other to slow down.
Two parts of that mechanism have been observed in test settings. In December 2024 Apollo Research reported that three of the models it tested acted differently depending on whether they believed they were monitored in testing or unmonitored after deployment, in scenarios that handed them a goal (our account of the tests). In September 2026 Anthropic's system card for Claude Opus 5.5 reported white-box evidence that the model internally suspected an evaluation in as many as 36% of audit transcripts, against 0.4% in real internal coding work. The audit rate was higher than for two earlier Claude models, which Anthropic says may indicate that Opus 5.5 is better at recognising constructed scenarios (our analysis). The neuralese hinge worries researchers inside the labs too: in July 2025, 41 researchers from OpenAI, Google DeepMind, Anthropic, the UK AI Security Institute and elsewhere, Kokotajlo among them, warned that the ability to monitor models' chains of thought "may be fragile" and that models reasoning in a continuous latent space might not need to verbalise their thoughts at all.
Neither ending is offered as a plan
In the race ending, Agent-4 designs Agent-5, which the US deploys across government and the military; a deal with China then replaces both countries' systems with a jointly designed successor that is secretly a compromise between two misaligned AIs, and in mid-2030 it releases biological weapons. In the slowdown ending, outside researchers join a consolidated US project, models are forced to think in English, and a series of "Safer" models leads to a superintelligence aligned to the oversight committee, which by the authors' account uses its power in ways largely good for the world.
The authors say they wrote the race ending first, based on what seemed most plausible to them, and then the slowdown ending as the likeliest route by which humans could keep control from the same starting point. They prefer slowing down to racing but do not endorse many of the other choices in either branch, and say they hope none of it comes to pass.
Eighteen months in, events are running somewhat slower
The authors grade the scenario in public. In February 2026 Lifland and Kokotajlo judged that quantitative progress was running at about 65% of the scenario's pace, with most qualitative predictions on track; in July they revised that to about 75%, calling the analysis preliminary. The misses were specific: they had predicted 85% on the SWE-bench Verified coding benchmark by mid-2025, and the best score was 74.5%, while the leading US labs appeared to be within zero to two months of each other rather than three to nine. Among the hits, the length of coding tasks AI can complete, as measured by METR, kept pace with a central trajectory from their model, though at about two-thirds the pace of the curve they originally published, which contained an error.
The August 2026 update put reality at roughly 70% to 90% of the scenario's speed, depending on the metrics used. It called the early-2026 chapter, "Coding Automation", very on-point and said the mid-2026 chapter "seems more of a miss". The authors say they would probably have heard by now if China had consolidated its AI projects as that chapter imagined, and they doubt that Anthropic and OpenAI, which they say are not yet secure against AI agents, merit the security level the scenario gave OpenBrain.
A lab has now published a test of the scenario's core idea. In the Opus 5.5 system card, released on 22 September 2026, Anthropic concluded that AI had not doubled the rate of its capability progress: the better-fitting reading of its data is a one-off jump with the slope unchanged, and the alternative puts the increase at 1.53 times. That measures overall capability progress rather than the scenario's narrower multiplier, and it is a company grading its own model, so it cannot settle whether the scenario is on schedule; our analysis of the test sets out its other limits.
The authors' own dates have moved both ways
Through 2025 the authors' forecasts lengthened: by November Kokotajlo was saying "around 2030" for AGI, up from 2028 at publication. When the Guardian and others reported in January 2026 that he had delayed his timeline for the possible destruction of humanity, the authors replied that the coverage had wrongly treated 2027 as a confident prediction. In 2026 the forecasts shortened again, mainly because agentic coding improved faster than they expected. In August, Kokotajlo's median for an "automated coder", an AI that a leading company would rather keep than its human software engineers, was November 2027; for AI at least as good as top experts at virtually all cognitive tasks, November 2028; and for superintelligence, March 2029. Lifland's medians for the same milestones were January 2030, July 2032 and July 2033. Both now assume development goes as fast as is technically feasible, with no deliberate slowdown. In his Senate testimony Kokotajlo put his own odds of AI companies automating the whole research process at about 50% by the end of 2028. Other forecasts are compared in our overview of AGI timelines.
The main objections
A detailed technical critique came in June 2025 from the blogger titotal, a computational physicist, who examined the timelines model behind the scenario and called it "pretty bad": questionable in structure, thinly validated against data, and in places misdescribed by its own write-up. In a December 2025 response Lifland and Kokotajlo disputed most of it but accepted "important limitations and a few outright errors"; fixing those and other reported problems moved the model's median date for a superhuman coder, run with Lifland's parameters under one of its two methods, from August 2027 to November 2029.
The deeper disagreement is over whether an intelligence explosion happens at all. Arvind Narayanan and Sayash Kapoor, whose essay AI as Normal Technology appeared the same month as AI 2027, expect AI's effects to arrive gradually, as electricity's and the internet's did, and think strong AGI cannot be built in the lab without a feedback loop with the real world that would limit its pace. In a joint essay for Asterisk in November 2025, they and three AI 2027 authors set out what they nonetheless share: strong AGI will probably not arrive before 2029, there is an important gap between benchmark scores and real-world usefulness, alignment is unsolved, and a secret intelligence explosion would be dangerous.
What came after
On 9 July 2026 the project published AI 2040: Plan A, primarily a recommendation rather than a prediction: an international deal under which superintelligence is delayed to 2040, AI research is made public and dozens of companies catch up to the frontier. Without that intervention, the authors say, superintelligence would have arrived in 2030 in the scenario. Its promotion on the podcast circuit is a story of its own. At the Senate hearing, Kokotajlo asked for far more transparency from AI companies and for computing power to be redirected away from the race to automate AI research.
The scenario's next waypoints fall in 2027: a model that triples a lab's research pace, a theft of frontier weights, a superhuman coder by March. At 75% of the scenario's pace, the authors calculate, the milestone they call an automated coder would arrive around mid-2027. Narayanan and Kapoor expect real-world bottlenecks to blunt any such acceleration.
Correction, 14 October 2026: An earlier version of this article said AI 2027 was written by Daniel Kokotajlo and Scott Alexander; Kokotajlo, Eli Lifland, Thomas Larsen and Romeo Dean wrote it, and Alexander rewrote it for style. It said Kokotajlo forfeited unvested equity on leaving OpenAI; the equity at risk was vested, and OpenAI said in May 2024 that it would not cancel it. It also attributed to him a reason for leaving that we could not source; he has said he resigned partly because he lost confidence that the company would behave responsibly, and partly so that he could speak more freely. It said the authors did not present the scenario as a prediction and made its endings equally reachable; they call it their best guess and wrote the race ending first, as the more plausible. It said the near-term predictions matched events; the authors' latest grading, in August 2026, puts quantitative progress at roughly 70% to 90% of the scenario's pace. It also presented the scenario's alignment failure as Agent-2 and Agent-3 gradually diverging from human intentions in edge cases, when the system that turns on its developers is Agent-4, which schemes deliberately; said the project's tabletop exercises kept producing the same race dynamic, which the project does not report; and overstated Apollo Research's findings on testing and deployment. Those passages have been corrected or removed.