AI Futures Project’s Plan A is less a prediction than a stress test for one hard claim: if superintelligence is coming soon, governance has to become a compute treaty, not a set of voluntary lab promises.
Source note: This analysis is based on AI Futures Project’s AI 2040: Plan A, available at ai-2040.com and as a PDF. The source is a scenario and policy recommendation, not an empirical research paper or neutral market report.
What This Is
AI 2040: Plan A is the sequel-shaped counterproposal to AI Futures Project’s earlier AI 2027 scenario. Where AI 2027 explored a rapid race toward superintelligence, AI 2040 asks what a deliberately better path would look like.
The report is explicit about its genre. Plan A is primarily a recommendation, not the authors’ best prediction. It is a scenario built to test whether their preferred policy package can survive contact with geopolitics, company incentives, labor disruption, covert projects, and alignment uncertainty. That matters because the document should not be read as a normal forecast. It is closer to a policy blueprint written as a future history.
The basic setup is stark: frontier AI companies continue racing toward systems smarter than humans in most economically and strategically important domains. The authors think this race creates two unacceptable outcomes. One is loss of human control. The other is a temporary monopoly over superintelligence by a tiny group of executives, officials, or states. Plan A is their attempt to avoid both outcomes without shutting AI development down forever.
The Core Thesis
The core thesis is that the world needs to replace a secretive race with a verified slowdown. In the report’s preferred path, humanity delays superintelligence until 2040, makes AI research transparent, allows many companies and countries to catch up near the frontier, and builds a regime where no state can safely defect by turning a large compute cluster into a sudden superintelligence project.
This goes beyond better safety culture. It is an international bargain around compute. Plan A assumes that the most dangerous AI progress depends on large datacenters, scarce high-end chips, and visible industrial infrastructure. If that is true, the policy surface includes chip tracking, datacenter monitoring, training-run controls, audit rights, and treaty enforcement.
The plan tries to solve two problems at once. First, it tries to buy time for alignment and control research to mature. Second, it tries to prevent any one lab or nation from holding exclusive control over superintelligent systems. The authors do not want one winner. They want a slower, more legible, more distributed climb.
The Argument Map
The scenario begins with a late-2020s wake-up period. AI systems become capable enough to disrupt labor markets, strengthen cyber operations, and convince policymakers that the race is no longer just a commercial competition. The United States and China face a choice between racing, sabotage, slowdown, shutdown, or a more complex deal.
Plan A chooses the deal. The first mechanism is compute visibility. Governments and auditors track major chip flows and large datacenters. The idea is not that every GPU on earth can be monitored perfectly, but that frontier training requires enough concentrated hardware to become visible.
The second mechanism is a pause or slowdown on frontier training. Existing AI systems can still be used for inference, product deployment, and safety work, but new dangerous training runs are constrained. Datacenters are monitored to prove they are not secretly training more capable models.
The third mechanism is total research transparency. AI research findings, algorithmic improvements, and safety results become public. This is the strangest and most important part of the plan. If breakthroughs cannot be hoarded, then the incentive to race for secret advantage weakens. Progress still happens, but it is supposed to become more predictable and evenly distributed.
The fourth mechanism is controlled growth. As AI and robotics become more economically powerful, the world limits the pace at which new compute and robots can be deployed. Governments auction permits for scarce automation capacity, then use the proceeds to fund a citizen dividend.
The fifth mechanism is mutually assured compute destruction. The report imagines large AI datacenters placed or structured so that treaty defection would not let one side simply seize a decisive compute advantage. If the deal collapses, the relevant compute can be disabled or destroyed before it becomes the basis for a unilateral intelligence explosion.
The Strongest Ideas
The strongest idea is that AI governance has to move from promises to verifiable physical constraints. A lab can promise to be cautious. A government can publish model rules. But if superintelligence depends on enormous compute clusters, the hard governance question becomes whether those clusters can be counted, audited, constrained, and disabled.
The second strong idea is total research transparency. Most AI policy debates assume that labs keep secrets and regulators inspect outcomes. Plan A flips that. It argues that algorithmic secrecy is what makes the race dangerous. If every new trick becomes public, a private lab gets less benefit from taking reckless risks. The frontier becomes less like a stealth weapons program and more like a regulated industrial sector where everyone can inspect the design logic.
The third strong idea is the link between automation rents and political stability. The report does not treat mass automation as a side issue. It assumes that if AI and robotics make human labor far less economically central, the legitimacy of the transition depends on who receives the gains. The citizen dividend is not a decorative welfare policy. It is part of the governance architecture. If the public absorbs the disruption while a few firms capture the upside, the deal breaks politically.
The fourth strong idea is scenario scrutiny itself. The authors do more than say coordination would be nice. They try to narrate the treaty, the enforcement problem, the economic transition, and the alternative plans. That makes the proposal easier to criticize, which is the point.
Load-Bearing Assumptions
The biggest assumption is that US-China coordination is possible under severe mistrust. Plan A requires both powers to accept invasive monitoring of their most strategic AI infrastructure. That is a huge ask. The report’s implicit bet is that fear of uncontrolled superintelligence eventually becomes stronger than the normal drive for national advantage.
The second assumption is that frontier AI remains compute-constrained enough to govern physically. If algorithmic breakthroughs allow small clusters to train dangerous systems, the compute treaty loses much of its force. The plan depends on the idea that covert projects are slower, weaker, and easier to detect than legal frontier projects.
The third assumption is that hardware verification can work. It is one thing to track chips in theory. It is another to prove, continuously and across jurisdictions, that a datacenter is only running allowed inference workloads rather than hidden training runs. The report treats verification as hard but tractable. That may be right, but it is not settled.
The fourth assumption is that alignment research benefits from a controlled climb to more capable systems. Plan A does not freeze everything forever. It allows slow scaling because the authors think powerful AI systems may help solve alignment and governance problems. If that belief is wrong, then Plan S, the shutdown path, may be safer than Plan A.
What Skeptics Would Challenge
Skeptics would start with political feasibility. A plan that asks frontier labs to give up secrecy, asks states to expose strategic compute, and asks the public to tolerate sweeping controls over AI development is not a normal regulatory program. It is closer to arms control plus industrial policy plus emergency economic redesign.
They would also challenge the confidence around covert projects. History gives many examples of smuggling, sanctions evasion, secret facilities, and state-backed deception. If a much smaller amount of compute can matter after years of public algorithmic progress, then defectors may not need a giant visible datacenter.
Economists would challenge the smoothness of the automation transition. The report imagines very large gains from AI and robotics, then tries to redistribute those gains through permits and dividends. But physical deployment is slow. Robots, power grids, factories, housing, healthcare, education, and regulation do not move at software speed. The political economy of replacing labor income with dividend income would be brutal.
AI researchers would challenge the boundary around total research transparency. What counts as AI research? What counts as ordinary software, mathematics, robotics, chip design, or product engineering? The more expansive the definition, the harder it is to enforce. The narrower the definition, the easier it is to route around.
What This Means for Builders
For builders, the useful lesson is not that Plan A will happen. It is that serious AI governance may reach deep into the build environment. The report points toward a world where frontier model work requires audit logs, safety cases, controlled datacenter access, model provenance, and verification infrastructure.
If research transparency became real, proprietary algorithmic tricks would become a weaker moat. Execution, reliability, product distribution, safety engineering, infrastructure efficiency, and compliance would matter more. The winning builder would not be the one who discovers a secret training recipe and races ahead. It would be the one who can build useful systems inside a transparent, inspected, tightly regulated frontier.
The report also makes safety cases feel more concrete. A safety case cannot be a PDF saying a model passed evaluations. In this world it becomes a live operational package: what model was trained, on what compute, under what constraints, with what monitoring, with what shutdown path, and with what evidence that the system remains under control.
What This Means for Buyers and Operators
For buyers and operators, the report suggests that frontier AI procurement could become more like buying critical infrastructure than buying SaaS. The key questions would not only be price, accuracy, and integration. They would include provenance, auditability, jurisdiction, permitted use, model lineage, and whether the vendor can prove compliance with international controls.
The citizen dividend sections also matter for operators because they highlight the labor transition. AI 2040 assumes that automation is not merely a productivity tool. It is a macroeconomic reallocation of work, capital, and political power. Any company planning around advanced AI should think beyond headcount savings. The deeper question is what happens when labor is no longer the main constraint on output.
There is also a warning for companies trying to build private AI advantage. In Plan A, secrecy is treated as a systemic risk. That does not mean all private AI work becomes impossible, but it does mean the most powerful systems would live in an environment where secrecy is suspect and verification is expected.
What to Read in the Original
The original is worth reading for the implementation details. The early sections explain why the authors think AI 2027 remains their default expectation. The 2029 to 2031 sections describe the political branch point and the formation of Plan A. The 2034 section on mutually assured compute destruction is the most distinctive governance mechanism. The 2035 pause at top expert AI explains why the authors do not want to scale straight to superintelligence. The later sections show the optimistic end state, including alignment becoming more scientific and humans gradually deciding when to defer to AI systems.
The alternative plans are also important. Plan B is sabotage. Plan C is a shorter slowdown. Plan D is the race. Plan S is shutdown. The comparison clarifies that Plan A is not the most cautious possible proposal. It is the authors’ attempt to balance safety, geopolitical feasibility, alignment progress, and the risk that a long shutdown eventually collapses into a worse race.
Bottom Line
AI 2040 is useful because it turns AI governance into an operational design problem. It does not stop at “slow down” or “regulate the labs.” It asks what has to be monitored, what has to be shared, what has to be capped, who receives the economic gains, and what happens if a treaty partner cheats.
The plan may be politically unrealistic. It may depend too heavily on compute remaining the main bottleneck. It may underestimate covert projects and overestimate international trust. But it is still a serious contribution because it makes the alternative to racing concrete. If AI 2027 is the warning, AI 2040 is the authors’ attempt to write the deal that would make the warning less likely to come true.
Source
AI Futures Project. AI 2040: Plan A. Available at ai-2040.com and as a PDF.