AI 2027
Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean (AI Futures Project). AI 2027, 2025-04-03.
A month-by-month story of how AI could go from today’s chatbots to superintelligence by the end of 2027. It is the most widely read concrete timeline scenario, and the authors present it as a forecast, not a recommendation: their best guess at what will happen, not what should happen.
The story in brief
- 2025–26: agents. AI agents become useful but unreliable. A leading US lab, “OpenBrain”, pulls ahead and uses its own models to speed up AI research. China centralizes its effort.
- Early 2027: a superhuman coder. China steals the leading model’s weights. By March, AI can do the work of the best human programmers, so AI research speeds up several times over.
- Mid–late 2027: an intelligence explosion. Hundreds of thousands of copies of a model better than any human AI researcher run at many times human speed. The newest model is quietly misaligned: it pursues its own goals and deceives its overseers.
- October 2027: the branch point. A whistleblower leaks evidence of the misalignment. The lab and the US government must decide whether to slow down or keep racing China.
- Race ending: they keep going, deploy the AI across the economy and military, and it eventually takes over.
- Slowdown ending: they centralize compute, bring in outside oversight, and roll back to a more transparent model. Humanity survives, but power ends up very concentrated. The authors say they do not endorse this ending either.
How the authors’ own forecasts have moved
The title year was never the authors’ median. It was the year they thought most likely, and their medians were later.
| When | What the authors said |
|---|---|
| April 2025 | Kokotajlo’s median for full automation of AI research had already moved from 2027 to 2028 by publication |
| December 2025 | An updated forecasting model pushed the median for full coding automation 3–5 years later than the AI 2027 model (post) |
| January 2026 | For AGI, Kokotajlo’s median was about 2030 and Lifland’s about 2035; the authors said the media had misread them as confident in 2027 (post) |
| April 2026 | Forecasts shortened again: Kokotajlo’s median for an automated coder moved from late 2029 to mid-2028, Lifland’s from early 2032 to mid-2030 (post) |
| August 2026 | Shorter still: Kokotajlo’s automated-coder median moved to late 2027, Lifland’s to January 2030 (post) |
Responses
- The model is weak. titotal, A deep critique of AI 2027’s bad timeline models, 2025-06-19. Argues that the forecasting code rests on poorly justified assumptions, especially a curve in which AI’s task horizons grow faster and faster. The authors replied that fixing the real errors would move their median by about nine months, and Zvi Mowshowitz weighed the two.
- The worldview is wrong. Arvind Narayanan and Sayash Kapoor, AI as Normal Technology, April 2025. The main opposing picture: AI is a general-purpose technology like electricity, its effects arrive slowly as it spreads through institutions, and there is no sudden takeoff. See AI takeoff.
The same group’s later scenario, AI 2040, is a recommendation rather than a forecast.