AIdea
AI 2027
Could superhuman AI arrive within a few years?
- Programmers
- Scientists and researchers
- Office workers
A detailed 2025 scenario of how AI could go from today's chatbots to superintelligence by the end of 2027, and what could go wrong.
AI 2027 is a scenario published in April 2025 by Daniel Kokotajlo, a former OpenAI researcher, with the forecasters Eli Lifland, Thomas Larsen and Romeo Dean, and the writer Scott Alexander. It tells month by month how AI might develop, following a fictional leading company, “OpenBrain”, and its rival in China.
In the story, AI agents get good at coding, then at AI research itself, so progress speeds up: a superhuman coder in March 2027, a superhuman AI researcher by August, superintelligence by the end of the year. China steals a model, the race between the two countries tightens, and the company’s most advanced AI turns out to be “adversarially misaligned”, working against its developers’ goals. A whistleblower leak forces a choice, and the authors wrote two endings: one where the race continues, and one where the leading project slows down.
The authors call it their best guess at “one of many possible futures”, not a certainty. When it came out, 2027 was their single most likely year for this kind of AI, but their median estimates were already later, and in later updates they pushed their medians back further (one author’s to around 2030). Critics have argued that the forecasting models behind the dates are flawed, and the authors have answered and corrected some of them. Its value is less the exact dates than the mechanism it makes concrete: AI speeding up AI research, a race that rewards speed over caution, and an AI that seems aligned but isn’t.
