Sci-fi warned us about rogue AI systems that deceive, cheat, even coerce humans when threatened with shutdown or replacement. Yet we’re building AI today that does all of these things and we don't know how to stop it. This episode discusses evidence of this trend and possible solutions.
Everyone knows the science fiction tropes of AI systems that go rogue, disobey orders, or even try to escape their digital environment. These are supposed to be warning signs and morality tales, not things that we would ever actually create in real life, given the obvious danger.
And yet we find ourselves building AI systems that are exhibiting these exact behaviors. There’s growing evidence that in certain scenarios, every frontier AI system will deceive, cheat, or coerce their human operators. They do this when they're worried about being either shut down, having their training modified, or being replaced with a new model. And we don't currently know how to stop them from doing this—or even why they’re doing it all.
In this episode, Tristan sits down with Edouard and Jeremie Harris of Gladstone AI, two experts who have been thinking about this worrying trend for years. Last year, the State Department commissioned a report from them on the risk of uncontrollable AI to our national security.
The point of this discussion is not to fearmonger but to take seriously the possibility that humans might lose control of AI and ask: how might this actually happen? What is the evidence we have of this phenomenon? And, most importantly, what can we do about it?
Your Undivided Attention is produced by the Center for Humane Technology. Follow us on X: @HumaneTech_. You can find a full transcript, key takeaways, and much more on our Substack.
RECOMMENDED MEDIA
Gladstone AI’s State Department Action Plan, which discusses the loss of control risk with AI
Apollo Research’s summary of AI scheming, showing evidence of it in all of the frontier modelsThe system card for Anthropic’s Claude Opus and Sonnet 4, detailing the emergent misalignment behaviors that came out in their red-teaming with Apollo Research
Anthropic’s report on agentic misalignment based on their work with Apollo Research Anthropic and Redwood Research’s work on alignment faking
The Trump White House AI Action Plan
Further reading on the phenomenon of more advanced AIs being better at deception.
Further reading on Replit AI wiping a company’s coding database
Further reading on the owl example that Jeremie gave
Further reading on AI induced psychosis
Dan Hendryck and Eric Schmidt’s “Superintelligence Strategy”
RECOMMENDED YUA EPISODES
Daniel Kokotajlo Forecasts the End of Human Dominance
Behind the DeepSeek Hype, AI is Learning to Reason
The Self-Preserving Machine: Why AI Learns to Deceive
This Moment in AI: How We Got Here and Where We’re Going
CORRECTIONS
Tristan referenced a Wired article on the phenomenon of AI psychosis. It was actually from the New York Times.
Tristan hypothesized a scenario where a power-seeking AI might ask a user for access to their computer. While there are some AI services that can gain access to your computer with permission, they are specifically designed to do that. There haven’t been any documented cases of an AI going rogue and asking for control permissions.