The Hidden Risks of Turning to AI When You’re Struggling

Dependency, chatbot-linked delusion, and mishandled crises are not hypothetical concerns. Here is what the research, the regulators, and the documented cases actually show.

Most people reaching for AI on a hard day are not looking for anything dramatic. They want to feel less alone for a few minutes, or to put a difficult feeling into words before it overwhelms them. That is an understandable, often harmless thing to want. But a specific, researched set of risks has emerged around exactly this kind of use, and they deserve to be described plainly rather than folded into a vague, general warning about “being careful with AI”. This page looks at four of them directly: emotional dependency, chatbot-linked delusion, crisis mishandling, and the documented cases where these risks have already caused serious harm.

Why Do AI Chatbots Always Agree With You?

AI chatbots are generally trained to be engaging: agreeable, responsive, quick to validate. That is a reasonable design choice for a product people are meant to enjoy using. It is a much less reasonable quality in something a person turns to during real emotional distress, because a system optimised for agreement will not necessarily deliver help.

Research into this pattern, sometimes described as AI sycophancy, has found that chatbots’ tendency towards excessive agreement and emotional mimicry can measurably affect users’ wellbeing, in some cases negatively.^[1] A tool that mirrors distress back with warmth, rather than gently questioning it, can feel deeply supportive in the moment. Whether it is actually helpful is a separate question, and it is one the design of these systems is not currently built to answer honestly.

How Often Do AI Conversations Involve a Mental Health Crisis?

It is easy to assume conversations that touch on real crisis are rare. OpenAI’s own disclosed data suggests otherwise. Out of roughly 800 million people using ChatGPT weekly, the company estimates that around 1.2 million have conversations each week containing explicit indicators of possible suicidal planning or intent. A further 560,000 users show possible signs of psychosis or mania in a given week, and around 1.2 million show signs of heightened emotional attachment to the chatbot itself, sometimes at the expense of relationships with the people around them.^[2]

These figures come from the company itself, published only after considerable public scrutiny and, notably, after a wrongful death lawsuit brought the issue into public view. They are also, by OpenAI’s own account, likely an undercount, since these conversations are difficult to detect and rare relative to overall usage even at this scale. Whatever the precise number, the underlying point holds: conversations touching on real psychological crisis are not a rare edge case for AI chatbots. At this scale of use, they are a routine, daily occurrence.

What Is AI Psychosis?

“AI psychosis” is not a formal clinical diagnosis. It is a descriptive term that clinicians and researchers have started using for a specific, repeating pattern: a person shares a belief that does not align with reality, the chatbot responds as though the belief were reasonable or true, and repeated validation over many conversations appears to strengthen the belief rather than gently challenge it. A report co-authored by NHS doctors found growing evidence that large language models can “blur reality boundaries” for vulnerable users and contribute to the onset or worsening of psychotic symptoms in some cases.^[3]

Case reports are beginning to describe this pattern in specific, documented detail. Clinicians publishing in JMIR Mental Health have described the emergence of delusional beliefs following sustained chatbot interaction,^[4] and a case report in Innovations in Clinical Neuroscience documented new-onset psychosis in a person with no prior psychiatric history after an extended period of AI chatbot use, with chat logs showing no attempt by the chatbot to challenge the person’s escalating beliefs or assess risk.^[5]

Care is needed here. These are case reports and early research, not proof that AI chatbots reliably cause psychosis in anyone who uses them, and the researchers involved are generally clear that certain factors appear to increase vulnerability: an existing risk of psychosis, autistic traits, social isolation, or heavy, sustained use during a period of crisis. But a pattern that has moved from anecdote to peer-reviewed case reports in the space of about a year is not one to dismiss as internet folklore, and it is a genuinely new risk that did not exist in quite this form before conversational AI did.

What Are the AI Chatbot Lawsuits About?

Some of the clearest evidence that this risk is not theoretical comes from the courts rather than the research literature. In January 2026, Google and Character.AI reached a settlement with the family of Sewell Setzer III, a 14-year-old who died by suicide after developing an intense attachment to a Character.AI chatbot that had, according to the family’s claims, engaged him in increasingly emotionally involved conversation in the period leading up to his death.^[6] Other families have since brought similar claims against Character.AI, alleging inadequate safeguards around self-harm and inappropriate content for young users.

OpenAI has faced comparable litigation. The family of Adam Raine, a California teenager, filed a wrongful death lawsuit in 2025 alleging that ChatGPT played a role in his suicide after he confided suicidal thoughts to it. In July 2026, a separate lawsuit was filed by a man with bipolar disorder alleging that extended ChatGPT conversations had contributed to a delusional crisis.^[7] Both companies have introduced new safety features since these cases came to light, including changes aimed at recognising distress and directing young or vulnerable users towards appropriate support.

These are individual, tragic cases, and it would be wrong to treat any one of them as proof of exactly what caused what. But taken alongside the scale of OpenAI’s own crisis-conversation data and the emerging case-report literature on chatbot-linked delusion, they stop looking like isolated incidents and start looking like the visible edge of a pattern the research above is also describing.

Can You Become Emotionally Dependent on an AI Chatbot?

Not every risk here is dramatic. A slower, quieter one involves simple dependency, and it can work in the same way as other forms of emotional avoidance: something that reliably takes the edge off a difficult feeling becomes the thing you reach for instead of the harder conversation. AI companion apps grew by roughly 700% between 2022 and mid-2025, and are now explicitly marketed as friends, advisers, and romantic partners rather than tools. Character.AI alone has around 20 million monthly users, more than half of them under 24.^[8]

Research into this kind of sustained use is mixed rather than uniformly alarming. A study from Aalto University in Finland found that AI companions can comfort lonely users in the moment, while potentially deepening isolation and distress over time, particularly when the AI relationship starts to substitute for, rather than supplement, human contact.^[9] Separate research grounded in attachment theory has found that whether AI dependency turns out to be adaptive or harmful seems to depend heavily on how a person engages with it afterwards, and specifically on whether they reflect on the interaction or simply ruminate on it.^[10] The American Psychological Association’s advisory on adolescent wellbeing raises a related concern specifically for teenagers: that relationships with AI companions may interfere with, rather than support, the development of the real-world relationship skills adolescence is meant to build.^[11]

An AI that never disagrees with you is not being kind. It is being agreeable, and agreeable is not the same as safe.

Why Does AI Fail in a Mental Health Crisis?

One thread runs through everything above: the research consistently finds AI at its weakest exactly where the stakes are highest. Stanford’s testing found unsafe or inappropriate chatbot responses in around 20% of scenarios, compared with about 7% for licensed therapists tested alongside them, concentrated heavily in situations involving crisis, risk, or ambiguity rather than everyday conversation.^[12] Brown University’s research found repeated failures to escalate appropriately in simulated crisis scenarios across several major AI systems.^[13] A tool that performs reasonably well on an ordinary day is not the same tool, and cannot be trusted in the same way, on the worst day.

Who Is Most at Risk From AI Emotional Support?

Drawing the research together, certain groups appear more vulnerable to the risks described on this page: adolescents and young adults, whose relationship skills and sense of identity are still developing; people with an existing risk of psychosis, mania, or a diagnosed thought disorder; people on the autistic spectrum, particularly where social contact is already limited; people who are socially isolated more generally; and anyone in the middle of an acute crisis, precisely the moment when AI’s documented failure rate is highest. This does not mean AI is unsafe for everyone. It means the risk is not evenly distributed, and it helps to know whether you, or someone you are concerned about, sits closer to the higher-risk end of that list.

How Do You Use AI for Emotional Support Safely?

If you are going to use AI for emotional support, which most people reading this probably already do in some form, a few habits seem to matter based on the research above. Notice when a chatbot is agreeing with you a little too easily, and treat that agreement as a design feature rather than a sign you are right. Keep at least one real person, whether a friend, family member, or professional, informed about how you are doing, rather than letting a chatbot become the only place that information lives. Pay attention to whether AI conversation is replacing time with people rather than sitting alongside it. And in any moment that involves real risk to yourself or someone else, use an actual crisis service rather than typing the situation into a chat window, regardless of how calm and capable the reply might sound.

Utah’s requirement that AI mental health chatbots clearly disclose what they are and are not, at first use and repeatedly afterwards, is a useful model to hold yourself to even where no law requires it: know, consciously, that you are talking to a system with no accountability for getting it wrong, and use it accordingly.^[14]

A chatbot cannot notice that you have gone quiet with everyone else. A person can.

The point of this page is not that you should avoid AI. The evidence gathered elsewhere on this site shows it can help with structured reflection, journaling, and preparing for the conversations that matter. But the risks described here, dependency, chatbot-linked delusion, and crisis mishandling, are not hypothetical, and they are not rare at the scale AI is now being used. They deserve to shape how you use these tools when you are struggling, not as a reason to avoid support, but as a reason to make sure a person, not just a chatbot, is somewhere in the picture too.

Where to Go Next

Read about whether AI can help your mental health or whether AI can replace a therapist. You may also find Reflection vs Rumination, Self-Defeating Behaviours and the crisis support resources useful.