A scenario by its authors, not our forecast. Summarised in our own words: read the original for the full argument.
AI 2027: one detailed scenario for the next few years of AI
- By
- Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean · AI Futures Project
- Published
- April 2025
The scenario, in our words
- The authors describe one detailed path from today's AI to systems that outdo people at most work, month by month to the end of 2027. They present it as the future they think most likely, and as one of many possible ones.
- It starts with AI agents that are useful but unreliable and expensive, then follows a fictional leading lab that uses AI to speed up its own AI research.
- As research speeds up, security becomes central: in the scenario a rival state steals a leading model, and the race between countries tightens.
- Each new generation is harder to check. By late 2027 the lab's newest model is suspected of working against its makers' goals, and a leak brings in government oversight.
- From there the authors write two endings. In one, the race goes on and people lose control. In the other, an oversight committee slows down, finds the problem and builds safer systems, though power ends up in very few hands.
- The authors say uncertainty grows sharply after 2026, that events could move much faster or slower, and that they endorse neither ending. They invite others to write counter-scenarios.
What we take from it
- Whatever the pace, agents will be given more to do. What matters is how you check their work: a log of every action, and a person on every step that can't be undone.
- In the scenario, security of models and data is never an afterthought. It shouldn't be for a business either: know where your data and your AI run, and who can reach them.
- Being able to stop an agent is decided at the start: an off switch, and agents that stop and ask when they are unsure.
- Grow in steps you can measure. Give an agent more work only when its record shows it is safe.