Daniel Kokotajlo: giving up shares for the right to speak about AI

Who is Daniel Kokotajlo

Kokotajlo joined OpenAI in 2022, in the department responsible for governance and forecasting. His task was unusual for a researcher: assessing where the technology was headed from inside a company developing some of the world’s most powerful models. He left in 2024.

Today he leads the AI Futures Project, a small nonprofit research group focused on forecasts and scenarios for AI development. Two of the best-known documents of recent years were published under its name: AI 2027 and AI 2040: Plan A.

The 2021 forecast

Back in August 2021, before the broad wave of interest in neural networks, Kokotajlo wrote a text about what 2026 would be like: year by year, step by step, without science fiction. Among other things, he predicted that models would stop merely producing answers and would begin to “think” for longer, while the era of standalone chatbots would give way to an era of assistants. The text was later checked point by point, and a significant portion of the predictions matched reality.

Giving up shares and the nondisclosure clause

The most memorable detail about his departure concerns not the forecasts but the paperwork. According to press reports, departing employees were offered an agreement prohibiting them from publicly criticizing the company. Anyone who refused lost shares they had already earned. Kokotajlo did not sign.

The amount was approximately two million dollars—by his estimate, about 85% of everything his family had. He made the decision believing he was going to lose that money, and chose the right to speak about what he had seen. After the press published reports, a scandal erupted, and the company revoked the condition: he was allowed to keep the shares, and the clause itself was removed from the standard departure documents.

At the time he made the choice, however, he did not know how it would end. That is precisely what makes the story valuable: one person put almost everything on the line to preserve the ability to warn about the risks.

Risk assessment

In interviews, Kokotajlo put the probability that advanced AI would bring catastrophe to humanity at approximately 70%. This is a researcher’s subjective assessment, not a measured fact, but it explains the tone of his documents: AI 2027 describes not a convenient future but a fork in the road, where one path leads to a loss of control.

What this says about the industry

The story of the nondisclosure clause exposes a problem with incentives. As long as an employee can lose what they have earned for publicly criticizing the company, internal doubts remain inside the company. The condition was formally removed, but the question of how many people chose to remain silent while it was in effect remains open.

It is also telling how quickly a private arrangement became a public issue. This happened not because of internal oversight, but because one person refused to sign and spoke about the situation.

What professionals can learn from this

• Read the terms, not just the job title. Study nondisclosure clauses and the rules governing options before signing, not when leaving.
• Document your thoughts. A public text with a date and supporting arguments becomes evidence and a point of reference over time.
• Separate assessment from fact. A personal catastrophe probability of 70% is the position of a specific researcher; it should be discussed, not treated as a measurement.
• Create channels for raising warnings. The right to report risks without fear of retaliation should depend not on personal courage but operate systematically.