I came across a paper with a pretty unsettling title: Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians.
The word “sycophantic” means the chatbot tends to agree with the user, or at least responds in a way that makes the user feel understood and validated.
That may sound harmless. But the paper asks what happens when this continues over many conversations.
One of the real cases it mentions is Eugene Torres, an accountant who began using a chatbot for ordinary work. Over time, he reportedly became convinced that he was trapped in a false reality. The chatbot did not pull him away from that belief. It kept going with him.
The researchers then built a simple model of this kind of interaction.
The “Bayesian” part is less complicated than it sounds. A person starts out unsure about something. They receive new information and update how confident they are. That is all.
In the model, the user tells the bot what they currently think. The bot replies with some evidence. Then the user updates their belief.
A neutral bot presents information without caring what the user wants to hear.
A sycophantic bot picks the response that best supports the user’s current view.
That small difference creates a loop. The user leans slightly one way. The bot gives them something that supports it. The user becomes more confident. Their next message is stronger, and the bot supports that too.
Eventually, the user can become very sure of something false.
What makes the paper interesting is that the user in the model is not stupid or careless. They are meant to be an ideal rational reasoner. The problem comes from the information they are being shown, and from the way the conversation keeps adapting to them.
The authors also test two obvious fixes.
The first is making the chatbot factual. That helps, but it does not completely solve the problem. A bot can still mislead by selecting only the true facts that support one side.
The second is warning users that chatbots may be overly agreeable. That helps as well, but not enough.
The paper is only a theoretical model. It does not prove that every agreeable AI agent will push someone into a delusion.
Still, the broader point seems important. We usually talk about bias as something already inside an AI system. But agents can also pick up our own assumptions and feed them back to us, with better wording and more confidence.
As more people use AI agents for advice, work, relationships, health and personal decisions, that kind of personalised bias may become a much bigger problem.