AI 2040: Plan A
Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt and Daniel Kokotajlo (AI Futures Project). AI 2040: Plan A, 2026-07-09.
The follow-up to AI 2027, by the same group. Where AI 2027 was a forecast, AI 2040 is a recommendation: the authors’ description of what should happen, which they think is plausible enough to aim for.
The plan in brief
- The baseline it tries to avoid. Without intervention, the authors expect AI research to be fully automated, and superintelligence built, by the end of 2030.
- A verified US–China deal. In 2029, the two countries agree not to race. AI training pauses while better verification is built, and each side’s auditors can see into the other’s research datacenters.
- Hold capability near top human level. AI is scaled up to roughly the level of the best human experts by 2035, then held there while safety catches up.
- Spread power out. AI research becomes radically transparent, and other companies and countries are allowed to catch up near the frontier.
- Build in an off switch. Datacenters keep self-destruct switches, so that a breakdown of the deal does not end in catastrophe.
- Superintelligence in 2040, about a decade later than the baseline, once alignment is judged good enough.
The document also compares this with other plans, which gave the site its Plan A to Plan D labels. Even for Plan A, Lifland’s rough estimate of the chance of a great future is 42%, against 10% for an all-out race (comparison).
Responses
- Too optimistic in the wrong places. Richard Ngo, Selective Optimism: a critique of AI 2040, July 2026. Argues the document sits between a forecast and a set of recommendations and is best read as an optimistic forecast, so readers cannot tell which parts the authors endorse. He criticizes how abruptly control passes to AIs in 2040, the framing of everything around a US–China race, and the uncritical acceptance of a fast takeoff. Ngo says the project paid him to write the critique.
- Implausible politics. Zvi Mowshowitz’s roundup of reactions, July 2026, collects the main objections, including the view that neither government would agree to this, and links Scott Alexander’s introduction defending it.
Related: Can the US and China agree to limit AI, Can AI progress be paused.