A former OpenAI employee proposed slowing down the AI race until 2040 - ForkLog

# Former OpenAI Employee Proposes Slowing the AI Race Until 2040

The authors of a forecast about the extinction of humanity due to AI have released a new scenario—this time with a chance of saving it. The AI 2040: Plan A document was presented by former OpenAI researcher Daniel Kokotajlo and the AI Futures Project.

Kokotajlo left OpenAI in April 2024. He left due to disagreements over artificial intelligence safety issues, and in 2025 he founded the AI Futures Project.

In April of the same year, the organization released the AI 2027 forecast. It said that the AI race between the United States and China could lead to the disappearance of humankind or to dictatorship. At the same time, the authors clarified that the probability of extinction ranges from 10% to 30%.

AI 2040: Plan A specialists called it a recommendation—described as a set of steps that can reduce risks.

How it should work

According to the scenario, in 2029 the United States and China sign an international agreement to refrain from racing for superintelligence. Without a deal, automation of AI development would have occurred as early as 2030.

Instead, the countries develop neural networks gradually—up to the level of the best human experts. By 2035 they pause to preserve human control over the systems, and in 2040 the pause is lifted, and AI reaches the level of superintelligence.

The plan is based on four principles:

• buy time for safety research;
• fully disclose AI development;
• distribute artificial intelligence among different companies and countries;
• keep the process reversible.

To ensure that the parties trust each other, Plan A relies on verification. Large data centers are visible from space, so they are difficult to hide.

This forms the basis of the first step: the countries publicly declare purchases of AI chips. Next, they introduce a temporary pause on new training runs—sensors at the data centers confirm compliance. Once trust is established, the restriction is lifted, but research remains completely transparent: any country and company can see what others are doing.

The deal should be protected from being derailed by “mutually guaranteed destruction of computing power,” which the authors compared to nuclear deterrence logic. Under the plan, new Chinese data centers are built in Canada, and U.S. facilities are built in Mongolia—on territory that, in the event of conflict, is easiest for the rival to attack.

If the agreement fails, the receiving country will try to seize the other side’s capacity on its own territory, and the owner will destroy it themselves so that it cannot be handed over to the opponent.

Economics of the plan

Based on the authors’ calculations, global computing capacity will grow from 20 million H100-equivalents in 2026 to 60 billion by 2034. Real U.S. GDP growth in certain periods of the 2030s will reach 50% per year—compared with the usual 3%.

Due to automation, employment in the United States will fall from 62% in 2027 to 12% by 2040. To compensate for job losses, the experts proposed introducing “civilian dividends”—payments to every adult American from income that the government collects from companies for permission to use computing and robots.

According to AI Futures Project forecasts, the dividend will be $45,000 per person in 2032. By 2035 it will rise to $1 million, and by 2039 to $10 million.

Four backup scenarios

The specialists set Plan A against four other possible paths of events. In their view, these are plotlines within a single scenario—an illustration of what would happen under other decisions by authorities.

• Plan B — The United States builds a coalition of allies and pressures China, up to cyberattacks and targeted strikes on the other side’s AI projects. The scenario ends with a choice between losing control over AI and war.
• Plan C — Washington tries to negotiate with Beijing and introduce internal regulation, but under pressure from companies, the pause is quickly lifted. The ending carries the risk of a permanent oligarchy of a small group that controls superintelligence.
• Plan D — Authorities bet on minimal regulation and a race. The risks are loss of control over AI, extreme concentration of power, and a third world war.
• Plan S — A complete indefinite halt to advanced AI development. The main danger is that the deal will fall apart sooner or later, and the race will resume under less controlled conditions.

Recall that in February, Citrini Research experts predicted the collapse of the economy due to artificial intelligence.