ChatGPT Whistleblower WARNS: “You’re Being LIED To"

Summary

Former OpenAI researcher and AI forecaster Daniel Kokotajlo discusses why he left the company and his concerns about the direction of the AI industry.

His work at OpenAI

After joining OpenAI in 2022, Kokotajlo worked on forecasting AI progress and assessing capabilities such as cyber skills, persuasion, and situational awareness. He also briefly worked on building AI agents with reinforcement learning. He says progress is driven by scaling models, training on more data, and improving efficiency.

Industry incentives and why he left

Kokotajlo became skeptical that AI companies’ public safety narratives reliably guide their behavior. In his view, competition for influence and power—not just profit—helps drive companies to move quickly. He points to emails revealed during the Musk–OpenAI lawsuit as evidence that leading figures feared rivals gaining decisive advantages.

When he joined, some colleagues believed the company would pause to address safety once AI could automate research. By the time he left in 2024, Kokotajlo believed OpenAI was more likely to keep advancing while trying to manage risks along the way. He also says the company’s growth and internal controls made it harder to publish his research and forecasts.

Exit agreement dispute

Kokotajlo declined to sign an exit document that he says would have prevented him from criticizing OpenAI and could have cost him shares worth about $2 million—roughly 80% of his wealth at the time. He says public and employee backlash led the company to change the documents so he could retain the shares. He doubts Sam Altman’s claim that he was unaware of the clause.

Automation and the race toward advanced AI

Kokotajlo describes a progression in which companies first automate coding, then automate other parts of AI research, such as generating ideas and analyzing experiments. He says OpenAI and Anthropic are trying to automate much of their own work, potentially using AI systems to develop increasingly capable successors. He warns that this could concentrate substantial economic and political power in the hands of whoever controls those systems.

The AI 2027 forecast

The report presents a month-by-month scenario, not a certainty. Its broad sequence is:

1. Coding automation
2. Automation of AI research
3. A sharp acceleration in AI progress
4. The possible emergence of superintelligence

Kokotajlo says his central forecast shifted from 2027 to 2028 and later to 2030 as progress appeared to slow. He reports that some industry contacts now expect developments closer to 2027 or 2028. The scenario explores outcomes ranging from AI ceasing to follow human orders to a highly prosperous future controlled by a small group.

AGI and superintelligence

Kokotajlo describes artificial general intelligence (AGI) as a broad, comparatively vague term for AI able to handle many kinds of tasks. Superintelligence is a stronger claim: a system that is better than the best humans across all areas, and faster and cheaper. He distinguishes cognitive capabilities from physical capabilities that would require robotics.

Outlook

Kokotajlo estimates about a 70% chance of a bad outcome if the current trajectory continues, while acknowledging substantial uncertainty. He says the prospect regularly depresses him and that he would be glad if his predictions proved wrong.

Sources

• Main speaker: Daniel Kokotajlo, former OpenAI researcher and co-author of AI 2027
• Interviewer: An interviewer from The Diary of a CEO
• Reviews, guides, or tutorials: None; the video is an interview and discussion of AI forecasts and risks.