Last year, the AI 2027 scenario spooked everyone with its forecast that AI would lead humanity to either extinction or dictatorship. Now the same team has released a follow-up — "Plan A" (AI-2040): their version of how to actually arrive at a good outcome.
The idea is counterintuitive. Instead of racing to "build superintelligence first," they propose deliberately hitting the brakes. To start, all AI research goes public — dozens of companies worldwide catch up to the leaders, and countries enter a nuclear-deterrence-style regime applied to compute (the authors call it mutually assured compute destruction).
What Plan A proposes
- By default, the researchers believe AI would have fully automated its own development by 2030 and produced superintelligence by year's end. Plan A delays that: until 2035, models are only grown to the level of top human experts, then paused — and the pause lifts only in 2040. Hence the project's name.
- The whole thing hinges on a US-China deal in 2029 to slow development down. Without it, the structure collapses.
- This isn't an attempt to predict the future. The authors describe what they think the right course of action would be — and admit themselves that reality will probably move faster. Daniel Kokotajlo (AI safety researcher, co-author of the AI 2027 scenario) already [thinks 2040 is too optimistic](https://blog.aifutures.org/p/q1-2026-timelines-update).
- "Plan A" is just one of five options. There are also Plans B, C, D, and S — each describing a different way the US might respond to approaching superintelligence. Or not respond at all.
Bottom line
A good ending here is only possible if the US and China actually agree to slow down together. If they don't — we get that same dystopian AI 2027 that this same team already mapped out.
All the scenarios are at ai-2040.com.