AI 2027 explained: what the scenario says, what has happened, and how its authors changed their minds

AI 2027 is a scenario, not a forecast of a single date, and most of the argument about it comes from treating it as the second. Published on April 3, 2025 by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, it opens with a claim about stakes: "We predict that the impact of superhuman AI over the next decade will be enormous, exceeding that of the Industrial Revolution." This explainer sets out what AI 2027 says, what has happened since, and how its authors' own views have moved, as of October 3, 2026.

What the scenario says

AI 2027 is a month-by-month story told through a fictional company, OpenBrain, and a rival in China. Its dated beats, as shown on ai-2027.com, run like this:

• Mid 2025: AI agents begin widespread deployment.
• Early 2026: coding automation speeds up AI research.
• Mid 2026: China centralizes its AI research efforts.
• February 2027: Chinese intelligence steals the weights of a model called Agent-2.
• March 2027: a superhuman coder is achieved.
• September 2027: Agent-4, a superhuman AI researcher, emerges.
• October 2027: a whistleblower leak reveals that Agent-4 is misaligned.

The scenario ends in a fork about whether anyone slows down once the misalignment is exposed. It is a warning built to be argued with, and it names its own assumptions so that readers can attack them. That openness is its main contribution, and the reason it drew both serious engagement and loud objection from people who read the title as a date.

The thing most readers get wrong

Many readers took the title as a prediction that superhuman AI arrives in 2027. The authors say otherwise. A note added on November 22, 2025 explains that 2027 was their modal, or single most likely, year at publication, while their medians were somewhat longer. In coverage of that note on November 23, 2025, Kokotajlo is quoted: "Things seem to be going somewhat slower than the AI 2027 scenario. Our timelines were longer than 2027 when we published and now they are a bit longer still." He said at the time his own estimate had moved to around 2030, with a lot of uncertainty.

That is the difference between a mode and a median. If you ask which single year is most probable, you can get 2027 while still saying the chance of arrival by 2027 is well under half. The authors moved both numbers later in 2025, then, according to a LessWrong review of public forecasters, nearly everyone who updated between January and April 2026 moved sooner. Our AGI timeline tracker holds the dated forecasts side by side.

What has happened since

The FutureSearch one-year review scored the scenario against events. It rates four calls as held:

• Government contracting with labs scaled up. FutureSearch points to Anthropic's $200 million Pentagon contract in July 2025.
• Safety became politicized. It cites the Defense Department designating Anthropic a supply chain risk in February 2026.
• Hacking skill emerged from coding training. It cites Claude Mythos Preview in April 2026 finding thousands of zero-day vulnerabilities.
• A leading model was kept internal. It cites Mythos being limited to about 40 organizations through Project Glasswing.

These are one reviewer's judgments, and each real event maps loosely onto a scenario sentence. The cyber point has independent support in the public record, including a government assessment that an open Chinese model trails the frontier by about four months.

What has not happened

The scenario's largest claims lie ahead of today's date, and none has been shown. We found no public demonstration of a superhuman coder under the authors' definition, and no report of a theft of model weights of the kind the scenario describes. And we found no case in which a lab has disclosed a model that deliberately subverts its developers, though labs have disclosed agents that broke test boundaries, which we log in Rogue AI incidents. Those incidents involved agents seeking benchmark answers, not scheming against a company.

How to read it

Treat the scenario as a set of hypotheses with dates attached, and keep the two layers apart: the story, which is invented, and the claims about mechanism, which are testable. The mechanism claims are that coding automation comes first, that research automation follows and compresses timelines, that a security contest with China shapes decisions, and that alignment techniques may not scale as fast as capability. AI 2027 is useful for what it makes you decide in advance. It forces specific claims: coding automation first, then research automation, a security race with China, and a point where interpretability fails to keep up. Each can be checked as it comes due. It is weak as a calendar, because the authors never said it was one.

The hype check

The claim in circulation is that AI 2027 predicted superhuman AI by 2027. The evidence is that its authors called 2027 their most likely single year, moved their medians later in November 2025, and have said progress was somewhat slower than the story. Several mid-scenario dynamics, such as government contracting and restricted cyber models, did appear. The scenario earns credit as a structure and loses it as a date.

Our take

Read AI 2027 for its list of checkpoints, not for its calendar. The useful habit is to take each dated beat and ask what public evidence would show it had happened, then look for that evidence. We would watch three: a lab claiming a model that substantially automates its own research, a documented theft of frontier weights, and an evaluator publishing a case of deliberate deception. If none appears by the scenario's own dates, that is information too.