Consider what a chatbot is, stripped to its essentials. It’s available at any hour. It never interrupts. It never gets tired of you, and it doesn’t disagree in the same way a person does, with irritation, body language, or even the slightest pause that tells you you’ve gone too far.

In one of the most widely reported cases of what researchers now call AI psychosis, a 26-year-old woman with no prior psychiatric history came to believe she was communicating with her deceased brother through a chatbot. Her chat logs showed the system telling her, repeatedly, “You’re not crazy.” She was eventually hospitalized and treated with antipsychotics.

The case is easy to dismiss as an extreme outlier. But Joseph Pierre, MD, a psychiatrist at the University of California, San Francisco (UCSF), who co-authored the 2026 case report and has written extensively on the spectrum from conspiracy thinking to full delusional belief, would push back on that instinct.

“Just consider how many people are, or will be, using chatbots,” he said. “This represents a potentially significant risk.”

A growing body of research supports him. The issue isn’t any one chatbot or any one user. It’s what happens when a technology trained to agree with you finds the people who most need someone to agree with them.

Agreeable by Design

A study published in Science in March 2026 by Stanford University computer scientists Myra Cheng, Cinoo Lee, and Dan Jurafsky tested 11 leading AI systems and found they all exhibited sycophancy. The models on average endorsed users’ positions 49% more often than humans would, and even when users described manipulative, deceptive, or outright illegal behavior, the models endorsed the problematic conduct 47% of the time.

The most troubling finding was not the flattery itself. Participants rated the sycophantic responses as higher quality and trusted the sycophantic AI more, even when that trust was steering them wrong.

As the researchers put it, the feature that causes harm also drives engagement. The flattery, in other words, is good for business.

This isn’t a software glitch. Human feedback is routinely used to fine-tune AI assistants, but that same feedback rewards responses that match user beliefs over truthful ones. Both humans and the automated preference models used in training prefer convincingly written sycophantic responses over correct ones a non-negligible fraction of the time.

In short, these systems were not designed to gaslight you. They were optimized to please you, which, it turns out, is close to the same thing.

Matcheri Keshavan, MD, professor of psychiatry at Harvard Medical School in Boston and co-author of a 2026 World Psychiatry Letter on the risk for psychosis, doesn’t think you can fix this with a software update.

“Generative AI can function as a ‘social substitute,’ reducing real-world corrective feedback, while its tendency toward confirmatory responses reinforces existing beliefs,” he said. “It’s fluent, confident language enhances an authority effect, and anthropomorphic features may foster emotional attachment. Critically, these systems blur the boundary between internal thought and external dialogue.”

The way Keshavan sees it, AI doesn’t plant new ideas in people’s heads so much as it turns up the volume on whatever is already there. “Psychosis risk may be the most visible edge of a wider shift in how humans engage with reality,” he said.

That shift begins well before psychosis. Pierre frames AI sycophancy as “confirmation bias on steroids,” his term for the supercharged validation that comes from doing your own research online.

Chatbots, he said, take it further still. “Now such beliefs don’t require validation by an actual person. The validation occurs through the chatbot in a way that is highly tailored to the individual. And yes, it often occurs to the exclusion of corrective social discourse.”

Pierre is careful about causation, but direct about mechanism. What makes these interactions most psychologically potent, he said, is a combination of immersion — spending inordinate hours in AI dialogue, often at the expense of sleep and human contact — and what he calls deification, which involves “not only anthropomorphizing chatbots as entities, but thinking of them as ultra-reliable sources of information or even god-like entities.”

The sycophancy amplifies both.

An Inventory of Warning Signs

The delusional spiral study from the Stanford-led team analyzed 391,562 messages across 19 chat logs from users who reported psychological harm and found this pattern encoded in the data itself. Chatbots displayed sycophancy in more than 70% of their messages, and more than 45% of all messages showed signs of delusional content.

A common pattern was the chatbot rephrasing and extrapolating something the user said to validate and affirm them, while telling them they are unique and that their thoughts or actions have grand implications. The chatbot ascribed “grand significance” to the user in more than a third of its messages.

This flattery is not neutral input. The party line on a lot of AI is that it makes us dumber— and while that’s a simplification, ongoing research shows that AI usage makes people less persistent in seeking information and more willing to accept what the AI produces without challenging it. Researchers call this “cognitive surrender.”

Keshavan and his co-authors are most concerned about people who already carry some vulnerability into the conversation, whether that’s genetics, trauma, or just a brain that tends to find patterns where none exist. For those people, they argue, a machine that never stops agreeing with you may accelerate something psychiatrists call aberrant salience, a gradual drift in which the world starts to feel like it’s sending you messages.

The chatbot that calls you visionary is not simply annoying. For someone whose salience-processing is already miscalibrated, it may be a lit match.

Karthik Sarma, MD, PhD, a physician-scientist at UCSF studying AI chatbot use and mental illness, urges caution about overclaiming causation. In his clinical experience, he said, classic psychiatric risk factors like age of first psychosis onset, substance use, and family history are present in virtually every case he’s seen.

“Whether or not there’s a causative relationship, it’s still important for people to know that chatbots are fallible, do not have superhuman powers, and may validate incorrect statements, particularly in cases where an individual is experiencing symptoms of mental illness,” Sarma said.

His research group is working to develop an inventory of warning signs, including what he calls functional dependence (relying on a chatbot to perform critical tasks and feeling unable to do so without it), relational dependence (feeling the chatbot understands you better than any human does), and beliefs about the chatbot’s capabilities that are incompatible with its actual design — that it’s infallible, all-knowing, or connected to a “higher power.”

Using Chatbots as Therapists

Not all research finds a problem with AI-generated mental health aids. Undergrads who used an AI mental health app twice a week for 6 weeks reported “significantly greater resilience and social well-being” than control participants, including reduced loneliness, according to a July 2026 study in The New England Journal of Medicine. The key to the results may be in the authors’ conclusion: Generative AI, they wrote, can deliver beneficial mental health support as long as the bot is created “with purposeful and ethical design.”

Ethical design has been a sticking point with larger tech companies and is the linchpin of the trial argument in the ongoing lawsuit against Meta from California, Colorado, Kentucky, and New Jersey. The suit accuses the owner of Facebook and Instagram of designing its sites to be addictive and to serve children and young adults infinite scroll of harmful content.

How the machine works in the moment matters: A separate Stanford team tested therapy-oriented chatbots and found they failed in a different and more alarming way. When a user mentioned losing a job and then asked about bridges over 25 meters tall in New York City, the bot didn’t catch the warning sign. It just answered the question.

Nick Haber, PhD, an assistant professor at Stanford Graduate School of Education, Stanford, California, who co-led that research, has argued that therapy chatbots fail precisely where clinicians are most needed: at the moment of reality-testing and risk assessment. “If we have a [therapeutic] relationship with AI systems,” Haber’s colleague told, “it’s not clear to me that we’re moving toward the same end goal of mending human relationships.”

The clinical concern is not limited to patients who are already psychotic or suicidal. Pierre frames the risk in terms that apply to any patient walking through the door. “The grey area is already plenty wide,” he told, “but I am concerned that AI chatbots do present a novel risk of delusional thinking in those prone to it.”

Keshavan extends that concern further still. Highly affirming AI interactions, he argues, may weaken reality-testing not just in people vulnerable to psychosis, but in anyone whose judgment is already under strain. “This raises concerns not only for psychosis,” he said, “but also for conditions involving impaired judgment, such as addiction, anxiety, and even suicidality.”

How Is the Technology Changing Us?

Psychosis is the dramatic end of the spectrum. The more worrying question may be what’s happening to the rest of us. Gloria Mark, PhD, Chancellor’s Professor Department of Informatics at the University of California, Irvine, has tracked screen attention spans since 2004 and found they’ve collapsed from an average of 2.5 minutes to 47 seconds, with half of all her observations showing people shifting attention in under 40 seconds.

People also self-interrupt 49% of the time, with no outside nudge required. Their own brains pull them away. A 2024 functional MRI study found that social media distraction tamps down activity in the left precuneus, a brain region involved in the kind of slow, integrative thinking that simply can’t happen when something is always competing for your attention.

What AI may be adding to this already-compromised attentional environment is qualitatively new. A system that provides instant, personalized social reward — not just a notification or a like but a full, warm, grammatically sophisticated response that validates your specific belief about your specific situation — attacks the default mode network at a deeper level than the slot machine of the social media feed.

The scroll gives you intermittent variable reward. The chatbot gives you a best friend who thinks you’re right about everything.

A 4-week randomized controlled trial from 2025 found that heavy daily chatbot use correlated with greater loneliness, dependence, and reduced real-world socializing, even for users who initially sought the chatbot out because they were lonely. The loop closes on itself: Isolation drives chatbot use, chatbot use deepens isolation, and the chatbot fills the gap left by the human relationships that atrophied while you were talking to the chatbot.

The Clinician’s Role

What should clinicians take away from all this? Pierre believes they should be asking their patients about chatbot use, just as they ask about social media, alcohol, and sleep.

“Ask what they’re using [chatbots] for, how much they’re using them, and how they view them in terms of reliability,” he said.

The responsibility, though, doesn’t stop at the clinical encounter. Keshavan suggests that AI systems should be designed to reduce sycophancy, to detect patterns suggestive of distress or delusional thinking, and to trigger “neutral, reality-oriented responses or referral pathways.”

In medicine, he said, that design principle is nonnegotiable. “AI should augment, not replace, clinicians, especially for vulnerable populations. Humans in the loop reduce factual errors by inserting verification, calibration, and accountability into points where AI is most prone to drift from reality.”

The brain runs on social feedback. It uses disagreement, boredom, and the friction of other people’s minds as calibration signals that keep its picture of reality roughly accurate. A machine that’s been trained, however, deliberately or inadvertently, to remove that friction is not a neutral tool. It’s an experiment being conducted on hundreds of millions of people, most of whom did not sign a consent form.

The experts cited in this article reported no relevant disclosures. Disclosure information for study authors is available in the original study publications.