Direct Answer
AI 2040: Plan A is a detailed policy recommendation wrapped in a future scenario. It is not a treaty, a law, a consensus forecast, or a promise that everyone will become a multimillionaire. The authors ask the United States and China to stop a race to superintelligence, account for frontier compute, permit reciprocal inspection, make advanced AI research unusually transparent, and resume scaling only under shared controls.
The extraordinary income number comes much later. The economics supplement assumes AI automates essentially all cognitive and physical labor, robots and compute scale at extraordinary rates, and governments capture the value of scarce deployment permits. It then redistributes most US permit revenue as a Citizen's Dividend.
That discrepancy does not make the report worthless. It demonstrates why scenario claims need a date, a model, a population, a unit, and a list of assumptions attached.
Watch the Explainer
Video credit: AI educator and entrepreneur Vaibhav Sisinty. Watch the original video and follow Vaibhav on X. The video includes a sponsored Higgsfield segment; its temporary promotion is not reproduced here.
Source-Status Map
| Claim | Status | What it actually means |
|---|---|---|
| AI 2040 exists and recommends Plan A. | Primary-source fact | The AI Futures Project published the report on 9 July 2026 and names six principal authors. |
| Plan A will be adopted by 2029. | Scenario condition | The authors do not call this their best forecast. The scenario assumes successful adoption so the proposal can be stress-tested. |
| Automated AI R&D would produce superintelligence around 2030. | Forecasting assumption | This is the report team's capability model, not an observed fact or scientific consensus. |
| Humanity should pause at top-human-expert AI in 2035. | Recommendation | The proposed pause is intended to preserve human control while alignment and governance work continue. |
| People receive $13M per year by 2040. | Video headline, source drift | The current economics supplement models around $10M per US person in 2040, in 2025 dollars, under aggressive assumptions. |
| AI CEOs acknowledge catastrophic risk. | Documented, but heterogeneous | Several leaders have made serious risk statements. Their definitions, probabilities, timelines, and preferred responses are not the same. |
| Plan A prevents catastrophe. | Conditional model conclusion | The result depends on coordination, verification, security, transparency, alignment, and economic assumptions all working well enough. |
What AI 2040 Actually Is
AI 2040: Plan A is the third major release from the nonprofit AI Futures Project, after AI 2027 and the AI Futures Model. The named authors are Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt, and Daniel Kokotajlo.
The team makes an unusually clear distinction: implementing Plan A is its recommendation; the downstream events are predictions conditional on that recommendation being implemented. The purpose is scenario scrutiny: write enough operational detail that hidden costs, contradictions, and failure modes become visible.
Kokotajlo's background matters, but it is not proof that the report is correct. He joined OpenAI as a governance researcher in 2022 and resigned in 2024. TIME reported that he refused a non-disparagement clause while believing he was giving up roughly $2 million in equity, before OpenAI later changed that policy. That history helps explain the report's emphasis on transparency and external scrutiny.
The Five Paths Are Policy Families, Not Movie Endings
| Path | Core move | What it buys | Main danger |
|---|---|---|---|
| Plan A | A verified international agreement, slow scaling, broad transparency, and distributed frontier projects. | More time, more safety work, less concentrated control, and a common verification regime. | Requires extraordinary US-China coordination, state capacity, security, and technical verification. |
| Plan B | Create a larger US lead through cyber, supply-chain, or potentially kinetic pressure, then spend part of that lead on safety. | More unilateral control over the pace. | War, escalation, secrecy, centralized power, and a temptation to race after creating the lead. |
| Plan C | A leading company or domestic regulator slows US development without a full international deal. | Some extra safety time without immediate treaty dependence. | Competitors may continue; the slowing actor may lose leverage or reverse course. |
| Plan D | Continue the frontier race near maximum speed with only a small safety allocation. | Maximum near-term capability and commercial momentum. | The report's default loss-of-control and concentration-of-power risks. |
| Plan S | Indefinitely halt frontier capability progress until specified conditions are met. | The largest safety margin and a simpler headline rule. | Enforcement, defection, stagnation, and losing the useful human-range AI that Plan A wants for safety research. |
The supplement also distinguishes Plan C+ and kinetic versus cyber versions of Plan B. Its numerical rankings are the authors' own rough estimates. They are useful for understanding the team's reasoning, not calibrated probabilities that settle the policy debate.
Plan A's 2029-2040 Timeline
- 2029: negotiate and pause frontier training. The US and China declare major compute holdings, audit supply-chain records, and allow reciprocal inspection. Existing approved models can continue serving inference.
- 2030-2035: scale inside the human range. Legal projects resume controlled research, share far more information, and aim for systems roughly as capable as top human experts.
- 2032: cap compute and robot growth. Cap-and-trade permits limit the modeled deployment boom to roughly six-month doubling times.
- 2035: pause at top-human-expert AI. The consortium stops before general superintelligence and concentrates on alignment, verification, security, governance, and public deliberation.
- 2040: unpause under a new governance system. The scenario assumes the world has built enough technical and institutional control to proceed.
This is a chosen scenario clock. AI 2027 used a faster modal year at publication; AI 2040 moves the default automated-R&D milestone to 2030 to reflect the team's uncertainty. The change is another reason not to read a year in a scenario title as a deadline.
The Compute Deal Lives or Dies on Verification
The proposal is more than "the US and China promise to be careful." It tries to make cheating observable and unprofitable. The verification supplement proposes:
- declarations and supply-chain audits for major compute holdings;
- hundreds of reciprocal inspectors during initial implementation;
- close tracking of new AI chips, memory, networking, manufacturing, and transport;
- inference-only retrofits for large datacenters, so approved models can run without new frontier training;
- reporting and random spot checks for smaller high-performance clusters;
- cold storage for chips awaiting verified research facilities;
- special economic zones for concentrated robot and hardware production;
- human audits, physical security, software monitoring, and intelligence work to detect covert projects.
One proposed inference check partially recomputes random workloads using approved model weights. The supplement also discusses zero-knowledge proofs, memory challenges, and hardware modification, while admitting that some options remain speculative or immature.
Three Operating Principles
1. Buy Time Without Freezing All Useful AI
Plan A does not propose stopping every model or taking current AI away from users. It tries to block the largest new training runs, keep approved inference available, and then resume controlled progress inside the human capability range. The wager is that powerful but still controllable AI can accelerate safety, science, forecasting, and governance.
2. Make Frontier Research Radically More Transparent
The report wants multiple projects in multiple countries working near the frontier, with results exposed to broader challenge. Transparency is meant to reduce a second risk beyond misalignment: a tiny group privately deciding how a superintelligent transition should work.
The tradeoff is real. Publication can help safety researchers catch errors, but it can also leak capabilities to covert projects. The assumptions supplement explicitly says Plan A biases toward transparency because the authors worry more about poor governance inside legal projects, while still calling for much stronger security.
3. Keep Defection Reversible
"Mutually assured compute destruction" is a deterrence concept. Participants preserve credible means to disable frontier compute if another party defects. The aim is to make a secret race less attractive without requiring perfect trust. It is also one of the proposal's most politically and operationally difficult elements.
Where the Citizen's Dividend Comes From
The economic model does not simply print money and divide it. It assumes AI and robots can perform almost every economically valuable task, but their deployment is capped. Because additional permits unlock enormous output, companies bid heavily for them. Government captures that scarcity rent and returns much of it to citizens.
| Model step | Assumed mechanism | Why it matters |
|---|---|---|
| Near-total automation | AI and robots can perform essentially all cognitive and physical tasks. | Creates the potential for explosive output without proportional human labor. |
| Deployment cap | Compute and robot quantities are allowed to grow roughly twice per year for part of the 2030s. | Makes permits scarce rather than allowing unrestricted replication. |
| Permit market | Companies bid for the right to deploy scarce productive capacity. | Converts part of the automation surplus into public revenue. |
| Dividend corporation | Each US citizen owns an equal share; the permit-revenue fraction rises from 25% in 2032 to 75% in 2035. | Separates basic income from wages and unequal stock ownership. |
| Price transformation | AI- and robot-produced goods become cheaper, while land and human-bottlenecked services become relatively expensive. | A large dollar dividend does not buy every scarce good in the same proportion. |
The current supplement projects about 90% average annual growth in real output from 2032 to 2037. It models the Citizen's Dividend at roughly $1 million per person in 2035 and around $10 million in 2040, measured in 2025 dollars. Those are outputs of the authors' model, not externally validated forecasts.
$13M, $10M, or Something Else?
The clean answer is: the video and current source differ. The video uses approximately $13 million. The source now says around $10 million. Possible reasons include an earlier model state, a different chart series, rounding, or later revisions. Without a versioned snapshot proving the $13 million input set, the current primary source should win.
There is a second correction. The economics supplement's concrete distribution mechanism is US-specific. It says the US begins redistribution to other countries, but that is not the same as proving that every adult worldwide receives the US per-person amount.
The Hard Assumptions Behind the Beautiful Outcome
| Assumption | Why it is load-bearing | What failure changes |
|---|---|---|
| Full cognitive automation arrives quickly. | The model's median automation timeline is far faster than mainstream economic forecasts. | Growth and permit value fall sharply if AI complements more tasks than it replaces. |
| Physical automation follows. | Robots must become capable, cheap, manufacturable, and able to reproduce industrial capacity. | Housing, logistics, energy, healthcare, and material abundance remain bottlenecked. |
| AI labor doubles roughly every 50-55 days before caps. | Explosive cognitive supply drives the modeled growth curve. | Longer doubling times produce a much smaller 2030s economy. |
| US-China coordination begins in time. | The scenario starts the deal before automated R&D makes small covert teams more powerful. | Later implementation is harder to verify and leaves less reaction time. |
| Covert compute remains detectable. | The consortium must keep secret projects too small to overtake legal projects. | A hidden project can defect, race, steal algorithms, or collapse trust. |
| Inference-only verification works. | Society wants access to existing AI while preventing prohibited training and experiments. | Governments must trust more, shut datacenters down, or accept higher risk. |
| Frontier security withstands nation states. | Weights, algorithms, monitors, and verification systems become prime targets. | Leaks can empower covert projects and invalidate shared pacing. |
| Permit rents are priced and captured well. | The dividend depends on governments owning a large share of the automation bottleneck. | Owners of compute and robots capture more surplus; the public dividend shrinks. |
| Political distribution remains legitimate. | Money alone does not prevent concentrated control over compute, land, security, or decision-making. | Abundance can coexist with dependency, surveillance, or authoritarian power. |
| Alignment becomes good enough by 2040. | The final unpause assumes humanity has a reliable way to maintain control. | The central catastrophe risk returns at a larger capability level. |
The authors do not hide all of this. Their assumptions supplement repeatedly marks verification difficulty, paradigm shifts, state capacity, covert projects, and post-superintelligence governance as uncertain. That candor is a strength. It does not make the uncertainties disappear.
Jobs, Prices, and Purpose
The report models 95% of actual economic tasks automated in 2035, including newly created tasks. That is a task-share claim, not the same metric as the share of people employed. The video separately describes a world where only around 12% remain in traditional jobs. Those figures should not be merged into one statistic.
The model also predicts a strange price system. AI labor becomes extremely cheap. Robot labor takes longer. Land, positional goods, and genuinely human services become relatively scarce. A multimillion-dollar dividend therefore does not mean every scarce apartment, location, or human relationship becomes trivial to buy.
There is also a non-economic question the report cannot solve with a transfer. Work provides status, rhythm, bargaining power, identity, and social contact. A society can be materially rich while leaving people politically dependent or psychologically unmoored. Any serious abundance plan needs institutions for agency, ownership, education, care, community, and democratic control, not only income replacement.
Are All AI CEOs Saying the Same Thing?
They share one important premise: advanced AI can create serious risks. They do not share one probability, one definition of catastrophe, or one policy plan.
- At an Axios AI summit, Anthropic CEO Dario Amodei put a 25% probability on the future going "really, really badly." That phrase is broader than literal extinction.
- Elon Musk said there was roughly a 10% to 20% chance AI "goes bad," while continuing to describe the likely outcome as positive.
- In a Lex Fridman interview, Google CEO Sundar Pichai declined to provide a number. He called the underlying risk high but argued that visible danger can mobilize humanity to respond.
- The AI 2040 team's own footnotes put literal AI-driven human-extinction estimates among its members between 10% and 30%.
These statements establish that catastrophic risk is not a fringe topic. They do not prove that catastrophe has a measured 10%, 20%, or 25% probability, that every leader means the same outcome, or that Plan A is the correct remedy. Subjective risk estimates are inputs to decisions, not experimental frequencies.
The Strongest Objections to Plan A
International Coordination May Be the Least Plausible Step
Plan A asks strategic competitors to disclose their most important infrastructure, accept inspections, constrain national champions, expose frontier research, and preserve credible retaliation against defection. Nuclear arms control offers partial analogies, but AI compute, software, model weights, and algorithms are more commercially distributed and easier to copy than fissile material.
Compute May Stop Being a Reliable Control Point
The plan assumes restricting compute meaningfully restricts dangerous capability. A major architectural breakthrough, extreme efficiency gain, distributed training method, or stronger use of consumer hardware could weaken that control point. The assumptions supplement explicitly treats large paradigm shifts as dangerous to the plan.
Transparency and Security Pull in Opposite Directions
External review can expose failures and distribute power. It can also publish capability advances that make covert projects faster. Plan A argues that both security and transparency can improve dramatically, but the right boundary will be contested task by task.
The Economic Model Is Far Outside the Mainstream
The economics supplement acknowledges that mainstream economists forecast much slower growth. It cites one expert survey averaging about 3.5% annual GDP growth from 2030 to 2050 in a rapid-AI scenario, while Plan A models roughly 90% annual real-output growth for part of the 2030s. The difference is not a rounding error; it comes from fundamentally different beliefs about full automation, substitution, replication, and bottlenecks.
A Safe Model Can Still Produce Unsafe Politics
Even if technical alignment succeeds, the consortium controls the world's most consequential infrastructure. Inspection, special economic zones, chip tracking, compute permits, and robot geolocation could create a surveillance state if checks and rights are weak. The report tries to distribute power, but institutional design remains as important as model behavior.
A Better Way to Read Any AI Scenario
- Label every sentence. Is it a fact, recommendation, conditional scenario, model output, company claim, or personal probability?
- Find the first divergence from today. Which unobserved event must happen before the rest of the story begins?
- Trace the mechanism. Do not accept "AI creates abundance." Ask what produces output, who owns it, how prices change, and how value reaches households.
- List load-bearing assumptions. Capability, hardware, energy, robotics, security, coordination, law, and legitimacy all matter.
- Check units and population. Per person is not per adult; the US is not the world; nominal dollars are not real purchasing power; tasks are not workers.
- Look for source drift. Interactive models change. Record the access date and prefer the current primary source over a repeated headline.
- Run the failure branches. What happens if the deal is late, a party cheats, verification fails, automation is slower, or public distribution is rejected?
That method neither accepts nor dismisses Plan A. It turns a dramatic story into a set of claims that can be debated and improved.
Video Chapters
| Time | Section | Reading note |
|---|---|---|
| 00:00 | $13 Million a Year, or the End of Humanity | Headline; compare with the current $10M US per-person source. |
| 01:12 | What the video covers | Overview of the report and its stakes. |
| 01:52 | Why the AI race is dangerous | The authors' diagnosis, not a settled consensus. |
| 02:24 | AI building AI | Automated R&D is a forecasted threshold. |
| 03:28 | Intelligence explosion | A modeled capability dynamic. |
| 03:45 | Testing obedience | The control and evaluation problem. |
| 04:20 | Why companies cannot slow alone | Race incentives and coordination. |
| 04:56 | Three bad endings | Misalignment, concentration, and conflict risks. |
| 06:26 | Plans D, C, B, S, and A | Policy families with different failure modes. |
| 07:28 | Sponsored segment | Higgsfield promotion; temporary offer omitted here. |
| 08:25 | What Plan A is | International agreement and controlled scaling. |
| 09:26 | Counting AI chips | Compute declaration and supply-chain tracking. |
| 10:44 | Inspectors | Reciprocal verification. |
| 11:37 | Building the consortium | Institutional expansion beyond two states. |
| 12:46 | Buy time | Slowdown without removing all useful AI. |
| 13:08 | Make research public | Transparency-security tradeoff. |
| 14:19 | Mutually assured compute destruction | Deterrence proposal, not deployed policy. |
| 15:45 | The 2035 rule | Pause at top-human-expert capability. |
| 16:50 | What happens to jobs | Conditional economic-model output. |
| 18:31 | How the dividend works | Permit rents and redistribution. |
| 18:58 | Only 12% still work | Do not confuse people employed with task share automated. |
| 19:34 | What work's disappearance costs | Purpose, status, and agency. |
| 20:06 | What would you choose? | Policy question, not prediction. |
Bottom Line
AI 2040 is valuable because it refuses to stop at "slow down," "win the race," or "share the wealth." It tries to specify inspections, datacenter rules, research access, capability ceilings, robot permits, dividends, and a route to restarting. That detail creates something concrete enough to criticize.
Its beautiful outcome is also extraordinarily conditional. The current $10 million dividend is not free money waiting in 2040. It sits at the end of a chain that requires near-total automation, explosive real growth, functioning permit markets, legitimate redistribution, successful international verification, robust security, and a solution to control before the final unpause.
Read Plan A as an ambitious stress test for AI governance. Do not read it as a personal financial forecast. The practical lesson today is simpler: insist on measurable claims, external scrutiny, reversible deployment, distributed oversight, and explicit ownership before capability outruns the institutions around it.
Sources
- AI 2040: Plan A - main report
- Economics of Plan A
- Comparing Possible Plans
- Verification Plan
- Plan A Assumptions
- AI 2040 authors, contributions, and changelog
- AI Futures Project announcement, 9 July 2026
- AI 2027 - original scenario and methodology notes
- TIME profile of Daniel Kokotajlo
- Axios report on Dario Amodei's risk estimate
- Bloomberg video of Elon Musk's risk estimate
- Lex Fridman transcript with Sundar Pichai on AI risk
This article is educational analysis, not financial, legal, or professional advice. AI 2040 is a recommendation and conditional scenario. Its economic outputs should not be used for personal financial planning.