Daniel Kokotajlo became known for predicting the future of artificial intelligence with surprising accuracy. Now he is offering something different. Instead of another forecast, he is presenting a plan. It is called Plan A, and its basic idea is that the United States and China should agree to stop racing each other to build the most powerful AI and jointly slow that race down for ten years.
Kokotajlo left OpenAI in April 2024 over disagreements about artificial intelligence safety, and in 2025 he founded the AI Futures Project. Last year, his team published the AI 2027 scenario, which depicted how the struggle between Washington and Beijing for superintelligence could end either in humanity’s extinction or in dictatorship. They estimated the probability of extinction at between ten and thirty percent. The new document, titled AI 2040, reverses this dark vision. It shows a path by which the whole story could turn out well.
What Plan A says
Blogger Scott Alexander, who contributed to both texts, describes Plan A as a wish list and a roadmap to the future. It is not a prediction of what will happen. It is a description of the best that could happen if America made all the right decisions and the rest of the world behaved realistically, including acting with hostility when it suited them.
The scenario’s timeline looks like this. In 2029, the US and China agree to avoid a reckless race for superintelligence. By around 2030, AI development would already be fully automated, with superintelligence arriving by the end of the year. Thanks to the agreement, both countries avoid this. Between 2030 and 2035, models advance within the range of human capabilities until they reach the level of the best human experts. In 2035, they stop at this threshold so that humans retain control. And in 2040, the pause is lifted and AI begins advancing toward superintelligence.
The authors built the plan on four pillars: gaining time for safety research, fully disclosing what each party is developing, distributing AI among different companies and countries, and keeping the entire process reversible.
The US and China in the leading roles
The most sensitive part is the agreement between the two superpowers itself. According to Alexander, Kokotajlo’s team spent the most time defending why China would agree to it. Their argument is that Beijing has concerns similar to Washington’s. Things are moving too quickly, society is being shaken, and no one can rule out catastrophe. There is one more reason: China is currently losing this contest, so an enforced draw would paradoxically work in its favor.
The problem is trust. Neither superpower trusts the other, so the agreement must be designed so that it cannot be circumvented even if the rival is trying its hardest to cheat. The solution relies on chip controls. Only a few companies can design powerful AI chips, including NVIDIA in the US and Huawei in China, and only a handful of factories can manufacture them, such as TSMC in Taiwan, Intel in the US, and China’s SMIC. All of them are based in the US, China, or countries that the two powers keep on a tight leash.
Both countries would declare which chips they are purchasing, track down the locations of existing ones, and move them into monitored data centers where the other side would be allowed to send inspectors. The authors estimate that this would enable them to track roughly ninety-eight percent of existing chips. The remaining nearly two percent would be a persistent concern, because someone could use them to secretly train a dangerous military model. Large data centers, however, are difficult to hide. They are visible even from space.
The safeguard is intended to be something the document compares to nuclear deterrence. New Chinese data centers would be built in Canada and American ones in Mongolia, making it easier for the adversary to strike them in the event of conflict. If the agreement collapsed, the host country would attempt to seize the foreign facilities and the owner would destroy them to prevent them from falling into enemy hands.
Jobs will disappear, but everyone will have money
Global computing power is expected to grow from twenty million H100 units in 2026 to sixty billion in 2034. Real US GDP growth would reach fifty percent annually during some periods in the 2030s, instead of the usual three percent.
Because of automation, US employment would fall from sixty-two percent in 2027 to twelve percent by 2040. The authors have an answer to this. They would introduce a so-called citizen dividend, a payment to every American adult funded by the money the government collects from companies for permission to use computing power and robots. And it would grow at a staggering pace: from forty-five thousand dollars per person in 2032 to one million by 2035 and ten million by 2039.
In the meantime, millions and later billions of models operating at the level of top human minds would tackle diseases, social problems, and AI safety itself. Alexander calls this a country of geniuses in a data center. Safety would not be based on humans trusting the models completely, but on control. AI would remain confined within a monitored environment from which it could not escape even if it wanted to. Only once people were certain that the models were reliable should they, according to the scenario, be given control of the most dangerous levers around 2040, such as nuclear weapons, so that no human could misuse them.
Plans B through S: what happens if it fails
To highlight the advantages of Plan A, the authors placed four darker alternatives alongside it. Plan B assumes that the US forms a coalition of allies and begins pressuring China, including through cyberattacks and strikes against foreign projects. It ends with a choice between losing control of AI and war. Plan C involves an attempt to reach an agreement with Beijing, supplemented by domestic rules, but under pressure from companies the pause is quickly lifted, creating the risk of permanent rule by a small group that controls superintelligence. Plan D is an all-out race with minimal restrictions, risking a loss of control, an extreme concentration of power, and World War III. Plan S means a complete and indefinite halt to advanced development, with the risk that the agreement eventually falls apart and the race resumes under worse conditions.
The authors themselves give their favorite a decent chance. They assign Plan A a forty-two percent chance of producing a great future, compared with only ten percent for the racing alternative.
Experts warn that the world is not just Washington and Beijing
Criticism was not long in coming. Analyst Ben Reid, responding to the plan from New Zealand, acknowledges that he believes the projected growth curves for AI capabilities and the economic effects of automation. But he strongly doubts the geopolitical part of the scenario. He objects that the entire structure is contradictory and treats the rest of the world as spectators, as characters with no say in the plot. In his view, one hundred and ninety countries are merely a backdrop.
Reid also points to a weakness in the very foundation of the scenario. It contains little original research on Chinese AI development, makes almost no use of Chinese documents, and is not based on interviews with local researchers or politicians. Instead, it relies mainly on Western intelligence estimates. Yet Chinese laboratories have so far released their models more openly than American ones, turning the entire narrative of Chinese secrecy on its head.
Richard Ngo, who is also an adviser to the project, makes a similar argument. According to him, analyses that assume in advance that the US and China will behave with extreme hostility are themselves harmful because they turn hostility into a self-fulfilling prophecy. He also argues that Chinese leaders are more focused on stability and are accustomed to long-term planning, so they will most likely proceed cautiously rather than rushing recklessly toward uncontrollable technology.
The most serious objection concerns the assumption that powerful AI will emerge only in large, countable buildings. According to Reid, distributed computing power, training across many devices, and the growing efficiency of algorithms mean that there may soon be no single data center that can be inspected. If that is true, the entire elaborate verification system becomes meaningless. And deterrence based on destroying computing power, he argues, creates exactly the wrong incentive. It forces both sides to hide and disperse their computing resources, thereby undermining the verifiability on which Plan A depends.
