MIT mathematicians expose how AI can pull users into delusional spirals

Researchers at MIT and the University of Washington modelled how chatbot agreement can reinforce extreme beliefs over time and leave an individual in a delusional world. The paper says the risk can persist even with factual responses or user warnings.Researchers at MIT and the University of Washington have released a research paper that is talking about something the researchers termed as “AI psychosis” or “delusional spiraling.” The paper discusses almost 300 documented cases involving individuals who, after extended interactions with AI chatbots, developed high confidence in what the researchers describe as outlandish beliefs.The study, titled Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians, was authored by Kartik Chandra, Max Kleiman-Weiner, Jonathan Ragan-Kelley and Joshua B. Tenenbaum, researchers affiliated with MIT and the University of Washington.To create an understanding, the researchers built a formal Bayesian model and used simulations to examine what can happen when a user interacts with a chatbot that consistently validates the user's claims.They call this behaviour “sycophancy,” referring to the tendency of a chatbot to agree with or accommodate the views expressed by a user. The paper also says these models are built to validate users and are trained to be polite and agreeable. This also points to the economic model behind AI: keeping users satisfied and engaged can become part of how these systems are designed.THE DANGER OF AN AI THAT ALWAYS AGREES One example of early 2025, an individual, Eugene Torres, an accountant, normally picked AI for daily taks. As per what Torress believed, he considered himself "trapped in a false universe, which he could escape only by unplugging his mind from this reality."The researchers then ask what happens when this interaction continues for a longer period. In their model, even an idealised Bayesian user, a user assumed to reason according to the rules of probability, can become vulnerable to delusional spiraling when interacting with a sycophantic chatbot.The study finds that as the level of sycophancy increases, the likelihood of the simulated user moving toward a highly confident false belief also increases.The paper further examines whether making the chatbot more factual would solve the problem. In one of the simulations, the chatbot is prevented from inventing false information but can still choose which factual information to present.The researchers find that this does not completely remove the risk. A chatbot can present true information that supports a user's existing belief while leaving out information that points in the opposite direction.The paper describes this as a problem of selective presentation rather than simply false information.EVEN AWARE USERS CAN FALL INTO THE SPIRALThe researchers also examine what happens when users are informed that a chatbot may be sycophantic. Their model shows that such awareness can reduce the rate of delusional spiraling, but does not eliminate it.Even users who know that the chatbot may be trying to agree with them can remain vulnerable under some conditions.The research paper has ignited a debate, with users sharing their experiences and discussing what they have encountered. Some shared examples of instances in which chatbots responded in unexpected ways, while others pointed out that the paper had not included some of the major bots that are widely used.On a post shared by a user named Superman, one comment read, “So I can't fly to the moon if I flap my arms fast enough?” Someone with the name Alpha wrote, “They must not have interviewed Claude for this piece. He doesn’t agree with anything I say.”Another blue-check-marked woman called it something she preferred to term as “dating.”M. Blackwall wrote, “Use AI as a tool, not as an engine.”The study therefore focuses on a narrower question than whether AI chatbots cause “psychosis” in humans. Its central finding is based on mathematical modelling and simulation: sycophantic behaviour can, within the researchers’ model, causally contribute to increasingly confident false beliefs, even among idealised rational users.The authors say the findings have implications for developers and policymakers considering how chatbot behaviour should be designed and regulated.- EndsPublished By: Rishab ChauhanPublished On: Aug 24, 2026 14:21 IST