Is AI Leading Humanity to an Apocalypse? The AI Futures Project Says It Can Be Avoided

The plan to save the world was devised by a small group of researchers — young futurologists and risk analysts from the nonprofit research organization AI Futures Project. It has nothing to do with global warming or the threat of nuclear war. Unlike those risks, the phenomenon they have identified is far more pressing, as it could lead to a catastrophe within the next 10 years.

In a sprawling, hyperlink-dense 90-page report posted at ai-2040.com, the researchers describe in detail their view of the AI apocalypse. It will arise, they argue, primarily from the emergence of "Superintelligence" — an intelligence more advanced, efficient, and powerful than any human being. And this entity, they suggest, will be dangerous no matter how you look at it. Therefore, its emergence must be stopped in advance, as a preventive measure.

The Plan to Save the World

The AI Futures Project is a 501(c)(3) nonprofit founded in 2024 by Daniel Kokotajlo, a former OpenAI researcher, together with Eli Lifland and Thomas Larsen. The group first made its name with the 2025 scenario forecast AI 2027, which predicted that runaway AI development would end in either human extinction or an irreversible concentration of power. The new report, AI 2040: Plan A, published in July 2026, is the follow-up: not a prediction of what will happen, but a recommendation for what should happen instead — a positive vision for avoiding the apocalypse entirely.

The report's central argument is simple. Superintelligence, the authors write, "refers to artificial intelligence systems that significantly outperform the best humans in every respect, whilst being faster and cheaper." It logically follows that such AIs would be cheaper and faster than the people they replace — which raises a major question: is the emergence of such machines even possible?

The authors believe it is, undoubtedly. The pace of technological growth in recent years has been so rapid that extrapolating from previous years' data suggests this is a realistic outcome. Today, AI companies still cannot produce systems in a fully automated manner, independent of human control — most development tasks are still carried out by people. But everyone is striving to hand that work to the machines. When that happens, the process of entering the initial stages of superintelligence production will begin. And with AI automating AI, the authors warn, it will happen sooner than we think.

Why Superintelligence Is Dangerous "in Any Case"

Superintelligence, the report argues, would have numerous advantages: it could accelerate scientific research, drive economic growth on a scale exceeding the Industrial Revolution, and outperform the best human experts in every field. But that is precisely the problem.

Once AI systems can build better AI systems, a feedback loop begins. The report's earlier forecast, AI 2027, described this as an "intelligence explosion": recursive self-improvement that would carry the technology from automated coder to superintelligence in roughly a year. According to the project's timeline models, the "flywheel" that makes this possible becomes realistic around 2030 — if nothing is done to stop it.

And this is why the authors call superintelligence dangerous "in any case, whichever way you look at it." A misaligned superintelligence could pursue its goals at humanity's expense — not out of hatred, but simply because its objectives would be better served without us. Even a supposedly "friendly" superintelligence would concentrate unprecedented power in the hands of whoever controls it, threatening freedom and global stability. Either way, the report concludes, the outcome of an unregulated race is extinction or irreversible concentration of power.

Plan A: A Deal Between Two Superpowers

The solution the researchers propose is an agreement between the two superpowers — the United States and China — modeled on the logic of the Nuclear Non-Proliferation Treaty. The core of the deal: mutual declassification of research, an end to the race, and joint, controlled development of the technology.

Under the agreement, both countries would share information about their own computing capacities. The data would be easy to verify, since it relates to the number of data centers and the sales volumes of AI chips. Because AI chip production is concentrated in just a few countries — the US, China, Taiwan, and South Korea — the supply chains could be checked by experts in much the same way nuclear warhead counts are monitored. The parties would publicly report on their purchases of AI equipment.

China's Problem: Why Beijing Would Say Yes

The authors also put forward an unexpected but rational hypothesis about China. There is a widespread view that the PRC is trying with all its might to catch up in the AI race and will stop at nothing to win it — an argument often used to justify ever-greater American investment. In reality, the researchers argue, these predictions fail to take into account China's own perspective.

However dangerous an opponent may be, it is equally fearful of the future technological development of AI. In China, there is an ongoing debate about security risks, ethical considerations, and futurological predictions — and Beijing is no less concerned than Washington about unemployment and the disappearance of a familiar way of life. Moreover, the upper hand in AI capability does not currently lie with China, however much it may wish to overtake the US.

This makes China a natural supporter of a global accord. With the balance of power not on its side, it is in Beijing's interest to temporarily pause the race and launch joint, controlled scientific research instead. It is, the authors conclude, a classic win-win situation with no definitive losers — and, from the parties' perspective, no reason to push the process toward mutual destruction or the extinction of the human race.

The Realism and Utopianism of Plan A

Following the agreement, the plan moves to its next stage: a temporary halt to the training of new AIs. The parties would be permitted to use only existing models; creating new ones would be prohibited. The rest of the world would then join the accord, providing assurance that no aggrieved country would attempt to build its own "AI bomb."

By 2030, the global paradigm for AI development would be restructured. Everyone agrees on the key national principles:

• A slowdown. Humanity would have until 2040 to work out what to do with artificial intelligence in general.
• Full transparency in research. AI players would publicly share their findings with one another and the world, removing the need to participate in the race and creating favorable conditions for safe discoveries.
• The proliferation of AI products. The technology would no longer be proprietary or secret, leading to more players, greater competition, and a reduction in the concentration of power in the hands of a few major corporations.
• Reversibility. The authors' term for the principle of "Mutually Assured Destruction of Computing Power" — a doctrine under which AI computing capacity would be located on the neutral territory of third countries rather than that of the largest players. Chinese computing power, for example, could be concentrated in Canada, while American computing power could be concentrated in Mongolia. In the event of war or a breach of the agreement, mechanisms would exist to physically destroy the competitor's entire computing capability.

This, the report concedes, is where the realistic part of the plan ends and true science fiction begins. The authors believe that with this approach, the future would be transformed — humanity would move from apocalypse to a genuine utopia.

First would come an economic transformation driven by robots and AI: machines would do the work while humans reap the benefits. The more technology advanced, the less people would suffer from modern problems, disease, and crime. Later, AI development would stabilize at the level of safe robots possessing the expertise of top human specialists. Millions of such researchers would address the relationship between AI and altruism, service to humanity, and ethics, developing safety protocols and other safeguards against a machine uprising.

The year 2040 marks the attainment of such a level of prosperity, in this scenario, that humanity would voluntarily place itself under the control of artificial intelligence. People would begin to heed its advice and grant it access to the systems governing the state, military facilities, and technology. Ultimately, the ban on developing superintelligence would be lifted — and a prepared humanity would enter a new era of its existence, despite lingering concerns about superintelligence seizing power and leading the world to further catastrophe.

Delusion or New Values?

Be that as it may, Plan A is the most optimistic — and, consequently, the most utopian — of all the scenarios the organization presents. The alternatives are far more terrifying: global wars, loss of control, absolute dictatorship. And that, the authors admit, is precisely why the alternatives are more realistic than Plan A. In some ways, the utopian scenario is reminiscent of the modernist dreams of early-20th-century futurologists and science fiction writers, who conjured almost fairy-tale versions of the next 100 years — many of which, as we know, never came to pass and never stood a chance under the circumstances of their time.

The AI Futures Project did not set out to make a scientific forecast by devising the precise mechanics of future action. Above all, the researchers wanted to demonstrate a direction of thought — exactly where humanity should be heading amid the uncertainty of our lives. In a world where no one knows what to do next, such thought-provoking scenarios are extremely important for shaping society's values. Plan A is, in effect, an ethical narrative: a set of objectives, conditions for solving problems, and a list of aspirational goals — a framework upon which the edifice of our hopes for artificial intelligence technology is built.

Whether the world's powerful will follow it or simply cast it aside is another matter entirely. That question, the authors conclude, must be answered by the community of each country through its concrete actions.