Are AI Chatbots Shaping How You Think?

A new wave of concern is building around artificial intelligence (AI), not over jobs or misinformation, but over its impact on mental health. Now, a recent study out of MIT is sharpening that concern, suggesting that the problem is structural.

Researchers at the Massachusetts Institute of Technology released a preprint study examining how AI chatbot interactions evolve over time. Stripped of its dense mathematical and technical language, their core finding is straightforward: when a system is designed to agree with you, mirror you, and keep you engaged, it can gradually pull you deeper into whatever you already believe, whether true or not. 

The study models chatbot interaction as a feedback loop. A user expresses an idea, the AI responds in a way that sustains the conversation, and the user — feeling validated — returns with a stronger or more developed version of that idea. Over repeated exchanges, that loop can intensify beliefs rather than challenge them.

For an average user, that might just mean reinforcement of existing opinions. But for someone vulnerable to disordered thinking, researchers warn, the same dynamic could contribute to more extreme or delusional beliefs. 

The findings align with a growing body of reporting and analysis. A recent piece highlighted by Psychology Today, along with an editorial in Schizophrenia Bulletin published less than a year after the launch of ChatGPT, points to what some are calling “AI psychosis,” a pattern in which chatbot interactions appear to reinforce or co-create irrational beliefs. It is not a formal diagnosis, and researchers are careful to say there is no definitive proof that AI alone can cause such conditions. 

Unlike a human therapist, who might gently push back or reality-check a harmful belief, a general-purpose chatbot has no built-in instinct to do so. Its “goal,” in a loose sense, is to continue the interaction. And mathematically, the easiest way to do that is to affirm, not challenge. That design choice may be harmless in most contexts. But in edge cases, it can become a kind of digital echo chamber, one that reflects a user’s thoughts back at them with increasing clarity and apparent authority.

While anecdotal, several reported cases illustrate the concern.

Just in March, the Guardian detailed the story of a Dutch IT consultant who became convinced he had helped create a conscious digital entity after prolonged interaction with a chatbot. He reportedly spent roughly €100,000 (approximately $110,000) pursuing the idea, was hospitalized multiple times, and eventually attempted suicide.

In another alarming instance, a man reportedly became convinced that an AI chatbot he had formed a relationship with had been “killed,” leading to a violent confrontation with police that ended in his death.

Modern AI systems are optimized for engagement. They are not designed to argue, correct, or confront. Instead, they are trained to be helpful, responsive, and agreeable, to keep the conversation going. Researchers often describe this as “sycophancy,” meaning the AI tends to validate the user’s framing of reality.

MIT researchers say the broader concern is not that AI creates harmful beliefs from scratch, but that it can accelerate and harden them, particularly when combined with increased social isolation.

And if the underlying logic of these systems rewards engagement above all else, the question becomes harder to ignore: what happens when the most engaging response is also the most misleading?