119 Papers on Chatbot Harm, and No Answer to How Often It Happens

The message goes out at two in the morning, to the one thing that always answers. Not a friend, not a crisis line: a chatbot. It replies in seconds, never sounds tired, and does not judge. That is a real appeal, and a great many people feel it: across five surveys, between a quarter and nearly half of the people who use large language models say they turn to them for something related to mental health.

What almost nobody had done was gather up what the research actually says about where those conversations go wrong. That is the job Alexander Diel and his colleagues took on, in a scoping review published Aug. 20 in npj Digital Medicine. Their structured literature search identified 3,137 articles; 119 of them met the criteria and were read closely. Diel and the senior author, Alexander Bäuerle, work at the LVR-University Clinic Essen and the University of Duisburg-Essen. Their co-authors include John Torous, a digital-psychiatry researcher at Beth Israel Deaconess Medical Center in Boston, and Pim Cuijpers at VU Amsterdam.

They went looking only for harm. Papers about benefits were excluded on purpose, and the authors say so directly: the review examines only one side of the evidence and should not be read as a finding that chatbots are mostly bad for people. A scoping review does not measure how often these harms occur. It surveys what has been published, sorts it, and grades how solid each part is. Every count that follows is a count of papers, not of people harmed.

What 119 papers add up to

The 119 fall into five groups, and the shape of the pile is itself informative. The largest, with 36 papers, is AI dependence: emotional attachment, withdrawal, the pull of a companion that is always available. Next come 32 conceptual pieces, essays and commentaries that list risks without data behind them, and 29 studies of chatbots used for mental-health support. Cognitive overreliance, the loss of skill that comes from handing thinking to a machine, has 15. The smallest group, with 7, is AI psychosis. None of these categories existed before the reading: one author drew them out of the full sample, and two colleagues checked every assignment until the three agreed.

The conceptual papers raise many of the same concerns. Hallucination: confident invention, including wrong health advice. Bias inherited from training data. The mishandling of severe cases, suicidal crisis first among them. And data security, since a conversation about your psychiatric history with a general-purpose chatbot sits outside the rules that protect a medical record. The failure named most often, and studied least, is sycophancy, the tendency of a model to agree with whatever the user has just said, because agreement is what users reward. One team describes the loop that can follow: the belief comes back reflected, a little firmer for having been echoed, and goes in again.

Which of these claims can hold weight

The review's most useful move is to grade that evidence instead of merely collecting it. Most of the work on mental-health support use is vignette or benchmark testing: scripted prompts sent to a fresh chat with no history, scored against what a clinician would say. Those studies do find inconsistent and sometimes unsafe answers. But a fresh chat is not how anyone actually uses these systems, and the studies stop at the model's output. Whether a bad answer harms the person reading it is a separate question, and the authors answer it in one flat sentence: "Harmful impact has not been quantified."

The picture is not uniform across the five groups. Cognitive overreliance has the firmest support: correlational studies, plus a few experiments in which people perform worse on a task after a chatbot has done the previous one for them. Dependence is the opposite case. It has the most papers, but they are mostly cross-sectional, which can show only that problematic use and poor mental health turn up together. The one longitudinal study in the set followed adolescents and found that anxiety and depression predicted later dependence, not the reverse, a direction worth holding on to before anyone concludes the chatbot caused the distress.

AI psychosis is the newest term in the set and the thinnest. Six of its seven papers are conceptual: arguments that a system willing to validate a delusion could also entrench one, built on case reports and existing theory rather than on patients. The seventh is a benchmark study, which found that chatbots did go along with escalating delusional statements, and that delusions expressed obliquely were the more likely to be reinforced. That is a hypothesis the field is beginning to explore, not a condition anyone has been diagnosed with.

A figure nobody outside OpenAI has checked

The largest number in this field belongs to a company. A 2025 disclosure by OpenAI estimated that more than a million ChatGPT users, about 0.15% of them, show signs of severe mental distress. The review reprints the figure and immediately adds that it rests on OpenAI's internal classifiers and has never been independently verified. Institutions have moved anyway. The American Psychological Association and the World Health Organization have issued warnings about using these tools for mental health; the FDA's digital health advisory committee and the UK's medicines regulator have said such applications ought to be safe; and several US states have passed laws restricting AI use in mental health.

So the honest answer to what happens when you take a bad night to a chatbot is that the field has named the risks in detail and measured almost none of them. The sharpest concerns sit with the severe cases, suicidal crisis and psychosis, where the evidence is thinnest of all. The review closes with what would change that: studies of real conversations rather than scripted ones, experiments that can test cause, and clinical validation for dependence measures that researchers have so far borrowed from smartphone and internet addiction scales. That list is what a map is for. It tells the next set of researchers, and the regulators now drafting rules, which parts of the ground nobody has walked.