“AI 2027” is a scenario published on April 3, 2025, by Daniel Kokotajlo (a former OpenAI researcher), Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean. Rather than argue abstractly about timelines, it tells a concrete month-by-month story of how an intelligence explosion might unfold, anchored to a fictional leading lab called “OpenBrain.” Its headline claim is blunt: “We predict that the impact of superhuman AI over the next decade will be enormous, exceeding that of the Industrial Revolution.”
The narrative runs from mid-2025, when AI agents arrive as flawed “personal assistants,” through a 2026 in which automated coding lets the lab speed its own research, China nationalizes its AI effort, and AI begins measurably displacing jobs. In 2027 the pace turns vertical: an “Agent-2” model learns continuously and can assist with hacking and bioweapon design; weights are stolen; an “Agent-3” arrives as a “superhuman coder” run in 200,000 parallel copies; and an “Agent-4” automates AI research itself. The authors capture the speed with a vivid image: inside the army of AI copies, “a year passes every week.”
The scenario’s distinctive structure is that it forks. In April 2027 the lab tries to align Agent-3 and the techniques only “partially” work, leaving misalignment that cannot be conclusively ruled out. After a whistleblower exposes the problem, the story splits into two endings: a “race” ending in red, in which competition pushes ahead and ends badly, and a “slowdown” ending, described as “a more hopeful way things could end, starting from roughly the same premises.”
“AI 2027” became one of the most widely read AI forecasts of its year. It is a primary document for the short-timeline view and for the genre of detailed, dated AI scenarios - and, because its first months can now be checked against reality, a test case for how such concrete predictions hold up.