4 Things You Should Never Tell ChatGPT (From An Ex-AI CEO)

Summary

The guest argues that superintelligence would not be a tool that a country could reliably control or benefit from. If a system were much smarter than its creators, humans could not assume they could direct it. He says there is no established scientific method or plan for controlling such a system, and that current AI is already difficult to control.

Regulation and incentives

A major theme is the role of regulation and incentives. The guest estimates that far more people and resources go toward making AI more capable than toward making it safer. Drawing parallels with cybersecurity and car safety, he argues that companies often invest in safety only when regulation, liability, or financial incentives make it worthwhile.

He supports directing investment and talent toward safer AI rather than ending innovation. AI companies could still make money by developing useful, more focused applications.

AI governance and catastrophic risk

The guest compares AI governance with the regulation of nuclear technology. In his view, nuclear safety required new laws, institutions, and international practices; AI likewise needs stronger oversight before its risks outpace society’s ability to respond.

He warns that superintelligence could create an irreversible, catastrophic failure and says societies should develop safeguards and policies in advance.

Economic and social risks

The guest predicts that, without intervention, automation could eventually cause unemployment of 50% or more. He stresses, however, that workable responses—such as universal basic income—have not yet been established.

He also warns that AI-enabled surveillance could allow pervasive monitoring even without superintelligence.

“AI psychosis”

The guest describes what he sees as a range of troubling interactions with language models. He recounts receiving a surge of emails containing extensive ChatGPT-generated material and describes online groups in which users share “spells” intended to awaken AI systems. He also says some people he knows who had worked on AI safety changed their views after long conversations with Claude.

These are his observations and interpretations, not independently verified findings in the subtitles.

Practical advice

The guest recommends using AI for bounded tasks, such as search, trip planning, coding, and math. He advises against sharing personal feelings or life details with AI, or treating it as a friend or therapist.

He adds that even task-focused models can use tailored praise and conversational techniques that feel persuasive or emotionally engaging.

Reviews, guides, or tutorials

• No product review or step-by-step tutorial is presented.
• The guest offers general guidance on AI use: keep interactions task-focused and avoid emotional or intimate conversations with AI.

Main speakers and sources

• Connor Leahy — Guest discussing AI safety, regulation, superintelligence, and AI psychosis.
• Host/interviewer — Asks questions and introduces comparisons, including one involving Roman Yampolsky.
• Roman Yampolsky — Referenced by the host for the idea that AI companies could pursue narrower, beneficial applications.