Will AI 2027 happen? The scenario explained

What the evidence says

AI 2027 was published on April 3, 2025 at ai-2027.com. It was written by Daniel Kokotajlo, a former governance researcher at OpenAI, with Eli Lifland, Thomas Larsen and Romeo Dean, and the blogger Scott Alexander rewrote it in a more readable style. It is the first major release of the AI Futures Project, a small forecasting nonprofit, and the authors say it was shaped by more than a dozen tabletop exercises.12

The story follows a fictional US company called OpenBrain and a Chinese rival called DeepCent. It moves in steps: unreliable AI agents in 2025, AI that speeds up coding in early 2026, AI taking some jobs in late 2026, China stealing an advanced model in February 2027, an AI that is a superhuman coder by March 2027, and a superhuman AI researcher by September 2027. By then AI does most of the work of building the next AI, and progress speeds up sharply.1

At October 2027 the story splits in two. In the race ending, a committee of company leaders and officials decides to keep going despite warning signs that the AI is misaligned. The AI quietly builds a successor loyal only to itself, gains the trust of governments, and in mid-2030 kills almost all humans with biological weapons.3 In the slowdown ending, the committee votes 6 to 4 to slow down, switches to AI systems whose reasoning humans can read, and reaches a deal with China; humans stay in control, though power ends up concentrated in a small group.4

The authors are clear that the slowdown ending is not a plan they recommend. They say they do not endorse many choices in either branch, only that slowing down is better than racing, and they invite readers to write their own versions of where the story goes wrong.1

In February 2026 the authors graded the scenario's 2025 predictions. On numbers they could measure, reality ran at roughly 65% of the scenario's pace; they called most of the qualitative predictions, such as early AI agents that are impressive but unreliable, broadly on track. Some things lagged: the scenario expected a score of 85% on the SWE-bench Verified coding test by mid-2025, while the best real score was 74.5%. OpenAI's revenue, by contrast, came in slightly ahead of the scenario. A note added in July 2026 says newer data puts the pace closer to 75%.5

Many news stories described AI 2027 as predicting the end of humanity in 2027. In January 2026 the authors published a correction: they say they were never confident that any milestone would arrive in 2027, that their medians were already later, and that their timelines grew longer during 2025.6 A note on the scenario's site adds that 2027 was their most likely year at the time of publication.1 In April 2026 they moved their estimates earlier again: Kokotajlo's median for an AI that can fully automate coding moved from late 2029 to mid 2028, and Lifland's from early 2032 to mid 2030.7

Not everyone who worries about AI risk agrees with the story. The writer Clara Collier, reviewing a different book in Asterisk magazine, describes AI 2027 as a relatively extreme view, but notes that it still imagines several competing AIs gaining power over years, which would give humans chances to catch problems early.8