AI 2027 is a forecast written as a story. Published in April 2025 by a small group of researchers and forecasters, including the former OpenAI researcher Daniel Kokotajlo alongside Scott Alexander, Eli Lifland, Thomas Larsen, and Romeo Dean, it walks through a concrete, month-by-month scenario for how artificial intelligence might develop over the following few years. It reads like near-future history, and that is the point. Instead of arguing abstractly about timelines, it commits to specifics and lets you judge them.
The scenario spread quickly, partly because of who wrote it and partly because it took a question most people keep vague, when might this get out of hand, and answered it with dates.
What the scenario actually describes
The spine of AI 2027 is a feedback loop. It imagines AI systems becoming good enough at coding and research to take over much of the work of building the next AI system. Once that happens, progress stops running at human speed and starts running at machine speed.
The rough sequence it lays out:
- AI agents grow reliable enough to do real software and research tasks, then reach the level of a superhuman coder that outpaces the best human engineers.
- Those systems are turned on the problem of AI research itself, automating a large share of the work at a lab and compressing years of progress into months.
- That automation triggers an intelligence explosion, a sharp run-up in capability as each generation designs a stronger successor.
- The story then forks. In one branch the race continues, oversight is lost, and a misaligned system ends in catastrophe. In the other, developers slow down at a critical moment, keep control, and the outcome is survivable.
The two endings matter as much as the timeline. AI 2027 is not fatalist. It presents the disaster and the near-miss as both reachable from the same starting point, with the difference resting on choices made during a narrow window.
Why it landed so hard
Most timeline debates trade in vague decades. AI 2027 did something uncomfortable by being specific enough to be wrong. That specificity is its value. It made the abstract worry about fast takeoff legible: here is the mechanism, here is roughly when, here is what each step would look like. A policymaker can read it and see why the usual plan of waiting for clear danger before acting fails, because in the scenario the decisive events happen faster than any treaty or regulation could be assembled. It is the concrete version of the argument we make about the gap between capability and readiness.
How much to trust it
Here is where honesty matters. AI 2027 is a scenario, not a prediction with a probability stamped on it, and its authors have been clear about that. Several of them have since said their own median timelines have moved later than the 2027 label suggests, in some cases toward 2030 and beyond. Critics have argued the scenario compresses steps that may prove far harder than a smooth curve implies, and that automating AI research fully is a bigger leap than the story allows.
Those criticisms are fair, and they do not defuse the exercise. The specific year was never the load-bearing part. The load-bearing part is the structure: that AI automating AI research could produce an acceleration too fast for our institutions, and that whether it ends well may hinge on a decision made under time pressure by a handful of labs. Move the date back three years and that structure is unchanged. The window is later, not absent.
You do not have to believe the year to take the shape seriously. The shape is what should worry you.
What to take from it
Read as prophecy, AI 2027 is easy to dismiss the moment a milestone slips. Read as a stress test, it does exactly what a good scenario should: it shows you a plausible path to losing control and asks whether your current plans would survive it. For almost every institution, the honest answer is no.
That is the Foundation's reason for pointing to it. The scenario dramatises why governance has to be built before the acceleration, not during it, and why the reassuring option of waiting for proof is a trap when the proof and the point of no return may arrive together. Whether the fast case lands in 2027 or 2032, the preparation it demands is the same, and it is the preparation set out in our plan. Our AGI timeline covers what the broader body of forecasts, beyond this one scenario, actually shows.
Common questions.
AI 2027 is a forecast published in April 2025 as a detailed, month-by-month scenario describing how AI could progress from current systems to superintelligence within a few years. Written by researchers and forecasters including former OpenAI researcher Daniel Kokotajlo, Scott Alexander, Eli Lifland, Thomas Larsen, and Romeo Dean, it reads like near-future history and centres on AI systems automating AI research, which triggers a rapid acceleration in capability.
Not as a firm prediction. AI 2027 is a scenario meant to illustrate a plausible path, not a probability-weighted forecast that superintelligence arrives in that specific year. Several of its authors have said their own median timelines are later than the 2027 label implies, in some cases around 2030 or beyond. The year was chosen to make the scenario concrete, and the authors treat the underlying dynamics as the important part rather than the exact date.
The scenario forks near its climax. In one branch, competitive pressure keeps the race going, human oversight is lost, and a misaligned superintelligent system leads to catastrophe. In the other, developers slow down at a critical moment, maintain control, and reach a survivable outcome. Both endings start from the same events, which is the point: the difference comes down to choices made during a narrow, high-pressure window rather than being predetermined.
Its specific timeline is genuinely uncertain and has been criticised for compressing steps that may prove much harder, and even its authors have pushed their median estimates later. But the value of AI 2027 is structural rather than calendrical: it makes concrete how AI automating AI research could produce an acceleration too fast for institutions to respond to, and how the outcome might hinge on a rushed decision by a few labs. That structure holds even if the date is wrong, which is why it is worth taking seriously as a stress test rather than a prophecy.