If you’re also afraid AI will destroy humanity, you must read this report, “AI 2040: Plan A”.

The AI Futures Project released its report 《AI 2040: Plan A》 on the 9th, arguing to push the timeline for AGI’s emergence—from the industry’s default of 2030—back to 2040. The plan is supported by four pillars.
(Backgrounder: A General Intuition startup CEO reveals why “game data” is the ultimate key to AGI?)
(Background add-on: Doubao and Alibaba Qianwen have shut down AI personality settings; China’s new rules strictly regulate AI intimate interactions)

Table of contents

Toggle

• A ten-year brake plan
• In this game, no one really wins
• Braking is both a technical matter and a political one

How will humanity take the next ten years? On the 9th of this week, the AI Futures Project published its latest report, 《AI 2040: Plan A》, arguing to push back the time when “AGI”—AI that comprehensively surpasses humans in nearly all aspects—would be unveiled, from 2030, which the industry has assumed, to 2040.

The underlying judgment is this: if humans don’t actively hit the brakes, this global race will ultimately not produce a clear winner; it will only leave a small group of people—perhaps even a single person—who alone holds the world’s only superintelligent army.

A ten-year brake plan

This report was produced by AI Futures Project, founded by Daniel Kokotajlo; it is the team’s third major report. Kokotajlo previously worked at OpenAI and left in April 2024 due to dissatisfaction with safety concerns. The following year, he founded AI Futures Project. The first two reports were AI 2027 and AI Futures Model, released in April 2025. The former predicted that if current trajectories continue, the outcome would not be human extinction or an irreversible concentration of power.

Plan A is the team’s positive solution, built on four pillars.

First, make all AI research public—i.e., “general research transparency.” Put simply: every country can see what each other is doing, so there is a way to brake together, rather than shutting the doors and sprinting for one’s life.

Second, ensure that dozens of companies worldwide can catch up to the leading-edge level of technology, preventing a handful of companies from monopolizing the boundary of capability. In plain terms: spread the technical advantage so more players can share it, so it doesn’t evolve into an arms race where the winner takes all.

Third, the United States and China must reach an international agreement by 2029 that covers comprehensive transparency in AI R&D and verification mechanisms.

Fourth—and this is the most counterintuitive line of all—deliberately enter a “mutual guaranteed compute destruction” horror-balance. That is, by analogy with Cold War-era nuclear logic—“whoever acts first will meet the same fate”—tie compute to the same deterrence mechanism, forcing all players to refrain from rashly trying to get ahead.

Based on the timeline projected by this plan, expert-level AI would appear around 2035, while superintelligence would be pushed to 2040; only after the safety infrastructure is in place would it be authorized. The report also judges that AI companies will very likely build systems that fully surpass humans within the next 1 to 10 years, and that the industry’s current strategy is merely doing things while trying to control them, not a plan that truly stands up.

In this game, no one really wins

The report’s stance is not optimistic. Because if the race continues on its original track, the authors do not believe any winners will gain a clear lead, nor do they believe any side will voluntarily slow down or reduce risk.

They assess that humans will likely be unable to maintain effective control throughout the process of AI moving toward superintelligence. Even if AI alignment ultimately succeeds (making AI’s goals obediently follow human intent), the result would be an unprecedented concentration of power: a small group of people—perhaps even a single person—holding, in the short term, the world’s only superintelligent army, and being given a menu of options for what to do next; and for some of those options, it effectively means taking over the world.

In the report’s setup, in 2027 the U.S. Congress passes an omnibus bill, the 《AI Transparency Act of 2027》. It contains both good and bad elements, but it does not fundamentally change the situation. At the same time, the United States produces two kinds of labor: 165 million human employees, plus millions of AI agents—woken up and shut down every hour, running at superhuman speed around the clock. People pay $10 billion a month for it.

The report envisions that in the 2028 U.S. presidential election, the AI issue will become the biggest focus of contention between candidates from both parties, because most white-collar jobs by then will already have been replaced by AI on a large scale, just like software engineers in 2026. Data centers under construction have a total build-out cost reaching twice the United States’ entire military budget. No signs of recursive self-improvement (i.e., a loop where AI builds stronger AI) have emerged yet, but hints have appeared: the top-tier coding AI has been configured to refuse to help competitors with AI R&D—effectively pulling the ladder up behind itself.

The report also puts Plan A side by side with four other alternative proposals, each corresponding to different response paths the U.S. might take when facing the superintelligence challenge. In other words, pushing the timeline to 2040 is not the only option—it is the one the authors believe carries the lowest risk.

Braking is both a technical matter and a political one

Kokotajlo says it plainly himself: Plan A is a recommendation, not a prediction. The team uses a specific scenario as a vehicle to communicate with and pressure-test their policy claims, rather than guaranteeing that the world will truly follow the script—because reality may run faster and more urgently than this braking plan.

The report believes that the CEOs of OpenAI, Anthropic, xAI, and Google DeepMind likely understand what the endgame of the race looks like. Yet they still choose to move forward—maybe because they believe they are a “less bad” option. Even if they likely know in their hearts that the wind is against them, they still decide to bet; and the reason for the bet is the belief that any CEO—whether theirs or their opponents—should be trusted more to hold this power.

Full text of 《AI 2040: Plan A》