Published July 9, 2026 by the AI Futures Project: Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt, and Daniel Kokotajlo, with transparency-plan and verification-plan supplements.
What it is
Plan A is the successor to AI 2027, which "predicted that this would result in either extinction or irreversible concentration of power." Plan A is offered as "our positive vision for what should happen instead."
The scenario's content in one sentence: "humanity delays the development of superintelligence until 2040, makes all AI research public, allows dozens of companies globally to catch up to the frontier, and intentionally enters a regime of mutually assured compute destruction."
The core mechanism is an international deal built on transparency rather than on capability limits alone: "The deal involves total research transparency for AI R&D, which allows the nations of the world to understand what's happening and enforce guardrails. The result is multiple companies across multiple countries scaling slowly and safely together towards superintelligence, instead of racing each other in secrecy."
The recommendation/prediction distinction
The authors are unusually explicit about the epistemic status of the document, and the distinction is load-bearing for how it should be cited: "Plan A is primarily a recommendation, not a prediction. This scenario is not our best guess as to what the future will actually look like. Instead, it's a vehicle for communicating and stress-testing our policy recommendations." The split is stated precisely — "while the implementation of Plan A is a recommendation and not what we actually expect to happen, the subsequent effects depicted are predictions."
Within the scenario, "Plan A is implemented successfully, albeit imperfectly and only in the nick of time."
Scenario scrutiny
The methodological argument is the document's most portable contribution. The authors hold that "most AI policy proposals fall apart under scenario scrutiny—that is, if you try to write down a detailed and plausible scenario in which that proposal succeeds, you will find it difficult to do so, and you will realize the plan is less likely to work than it seemed, or has more unpleasant side-effects than its proponents acknowledged."
Their explanation for its rarity is an incentive account: "applying scenario scrutiny to their own favorite policies might surface uncomfortable issues with them; meanwhile, applying scenario scrutiny to their rival's policies is a lot of work for little rhetorical gain." They apply it to their own proposal "even though this opens us up to criticism," and ask to be judged "against the existing state-of-the-art for plans to navigate the AI transition (if they can find any) and not against some hazy but pleasant fantasy where no one has to make any hard choices yet everything will probably be fine."
Against the objection that the exercise is unpredictable — "like trying to predict how to best fight World War 3, except that it's an even larger departure from past case-studies" — they cite precedent in military Taiwan gaming, intelligence agencies, climate bodies, and pandemic-preparedness offices.
The threat model
Two distinct failure modes are argued, and the second does not depend on alignment failing.
On control: the industry "has convinced itself that controlling superintelligent AI can be figured out on the fly, and thus has no remotely adequate plan." The authors "do not expect whoever 'wins the race' to have much of a lead," nor "to unilaterally slow down to reduce existential risk," and conclude that if the race continues they "do not expect humans to maintain effective control."
On concentration: "even if the AI companies somehow align their AIs, the result will be an unprecedented concentration of power—that is, the result will be a situation where a tiny group of people, or possibly just a single individual, is effectively in control of the world's only army of superintelligences for some months, and will be presented by said superintelligences with various options for how to proceed, some of which will de facto amount to taking over the world."
Their reading of frontier-lab leadership is that the CEOs of OpenAI, Anthropic, xAI, and Google DeepMind "understand this and are proceeding anyway, perhaps because they think they are the lesser evil and will use their immense power responsibly." The authors' response distinguishes the choice from the advocacy: "while we agree that it is generally correct to choose the lesser evil, we don't think we should advocate for a strategy that has such a scarily high chance of leading to human extinction or global dictatorship."
Timeline
| Year | Event |
|---|---|
| 2029 | The US and China agree to avoid a reckless race to superintelligence |
| 2030 | Fully automated AI R&D would otherwise arrive, leading to superintelligence by year end; the deal averts this |
| 2030–2035 | Scaling within the human range, to AIs "roughly as capable as top human experts" |
| 2035 | A pause at top-human-expert level "in order to maintain human control" |
| 2040 | Unpause and scale to superintelligence |
The default timeline moved from 2027 to 2030 to reflect a change of authorship rather than a change of view: AI 2027's year reflected Kokotajlo's roughly 50% median, while 2030 is Larsen's. The authors note that "Daniel currently thinks things will probably go somewhat faster than depicted in this scenario."
The five plans
At the 2029 decision point the scenario sets out five options, named for the strategic posture each represents:
| Plan | Posture |
|---|---|
| A | Verified Slowdown |
| B | Fight China |
| C | Burn the Lead |
| D | Race to ASI |
| S | Shut it all down |
A companion supplement, Comparing Possible Plans, scores these against explicit metrics — takeoff length from Automated Coder to takeover-capable AI (6 years under Plan A against 1.02 years under Plan D), safety compute (100B H100e-years against 500K), safety-researcher-years (1M against 200), three subjective 1–5 metrics, and probabilities of alignment (72% against 25%) and of a great future (42% against 10%) — and extends the taxonomy with Plan B-Kinetic, Plan B-Cyber, Plan C+, Plan E, Domestic-First Plan A, GPU arms control and CERN for AI. It also records per-author estimates of each plan's likelihood, placing Plan A itself at roughly 5%.
The incremental wishlist
Alongside the headline proposal, the authors set out less ambitious measures "that still help" — the part of the document most usable independently of the scenario.
Transparency. "The most important transparency intervention is limiting the gap between internal and external deployment," on the reasoning that "the internally deployed AIs are where most of the AI takeover risk comes from because those are AIs involved in recursive self-improvement," while external deployment "allow[s] the broader public to interact with and understand AI capabilities, which is vastly more informative than any abstract report or evaluation." Companies should also publish model specifications, information on whether models follow them, internal usage statistics such as "fraction of compute spent on internal deployment," and qualitative impressions of internal use.
Export-control enforcement, stated with an explicit reservation: "We have major reservations about introducing new export controls because they exacerbate the US/China race, but given the existence of export controls, we should obviously enforce them. If we don't enforce them, then we should consider repealing them." The concern is verification rather than diffusion — citing Epoch's estimate that "roughly a third of Chinese total compute is acquired via smuggling," smuggled chips "make future agreements based on compute governance more difficult to enforce."
Verification R&D. Not strictly necessary for an agreement but "extremely helpful": an "inference-only verification solution would enable the US and China to agree to stop doing new frontier AI training runs while allowing the public to maintain access to existing AI models."
Limiting AI R&D compute budgets. Noting that in 2026 "big AI companies spend roughly half their compute budget on AI R&D," capping that fraction "would slow capabilities progress, giving the world a bit more time to react."
Compute tracking, including directing companies "to stop recycling AI chips, because decommissioned chips are one of the most promising routes for covert projects to acquire chips later."
Government AI capacity: "The US government has barely any top-tier AI talent right now, so fixing this should be an urgent priority."
Relationships
- related: AI 2027 — the predecessor scenario this is written as the positive counterpart to
- related: Compute Governance — the verification and compute-tracking measures the deal depends on
- related: AI Power Concentration — argues concentration follows even if alignment succeeds
- related: AI Futures Project, Daniel Kokotajlo, Recursive Self-Improvement (RSI), AI Existential Risk, US-China AI Competition: Different Races, Different Metrics