AI 2040 Plan A: a recommended slowdown deal, not a prediction

AI 2040 Plan A is a detailed recommendation scenario for a verified US–China slowdown, research transparency, and compute tracking that would push generally superhuman AI toward 2040. It is not the authors’ best guess of what will happen, and it is not a lab result from this project.

All related news

What decision changes?

Treat Plan A as a concrete wishlist of outer handles—verification, shared evidence, release constraints—and ask which ones become independently usable under competitive pressure. Do not treat its dates or economics explorers as confirming this project’s micro simulations.

The AI 2027 team’s positive plan: delay superintelligence with verification and transparency—while still racing the hard technical problems.

AI 2040: Plan A is a long scenario from the AI Futures Project — the same group behind AI 2027. Unlike AI 2027, which was mainly a forecast of a fast race, Plan A is mainly a recommendation: what the authors think the world should try to do.

In short, they want the United States and China to make a deal that slows the rush to superintelligence, makes frontier AI research much more public, tracks major compute, and uses something like mutually assured compute destruction so cheating is hard. In their story, that deal delays generally superhuman AI until around 2040. They contrast this “verified slowdown” with racing, burning a lead, fighting, or shutting everything down. They say Plan A is not their best guess of what will happen — it is a detailed stress-test of a plan they think is worth aiming for.

This card is about that outer plan, not a new lab result from this project.

How it connects to earlier news on this site

The Pacing the Frontier statement asked for tools so companies and countries can slow automated AI development without losing alone. Plan A is a concrete (and contested) proposal for what those tools and deals might look like.
The AI 2027 lab stress-test used AI 2027’s speed assumptions only as schedule cues in a separate toy lab. Plan A’s takeoff and economics explorers sit in that same macro schedule family. They do not confirm or refute the lab findings, and the lab findings do not validate Plan A’s dates or geopolitics.
The secret-loyalties pipeline test is a reminder that strong visibility or research-sharing regimes still need directional checks inside the workflows that give a system power. Transparency alone is not correction.
The insurance-and-audits card asks who pays when a safety assessment is wrong — the same spirit as asking which Plan A handles are independently usable under pressure, not only persuasive on the page.

Reactions worth reading

Zvi’s overview walks through common objections (timelines, authoritarianism, China deal feasibility, economic optimism) without endorsing the plan as written. Nate Soares calls it the best concrete positive vision he has seen spelled out, while doubting it would work as written and treating a competent attempt as a possible off-ramp toward a harder shutdown—the hard-pause and off-switch priority in MIRI’s 2024 strategy update.

Questions to ask next

From the coordination and selection chapters: private caution does not become collective action by itself (Ch. 13, Ch. 34). From the attractor chapters: a pause buys time only if safety-relevant artifacts still reach release gates (Ch. 37–Ch. 38). From Appendix C: verification, compute access, and licensing are institutional handles — they must bind an uncaptured corrector when competitive pressure rises.

Plan A does not close this project’s open technical bridges. It is a serious field proposal for changing the selection environment in which those bridges would have to be earned.