A scoping review in npj Digital Medicine catalogs how LLM-based chatbots can harm mental health, drawing on 119 studies screened from more than 3,000 records.
The authors, led by researchers in Germany and the US including digital psychiatry expert John Torous, sorted the literature into five harm categories.
The conceptual arm of the literature points to chatbot weaknesses like hallucinations, sycophancy, and bias, plus data-security concerns and exposure for severe psychiatric conditions such as suicide and psychosis. In vignette-style experiments, the chatbots gave responses to mental health questions that fell short of clinical expectations.
Other strands link cognitive overreliance on chatbots to declines in academic and cognitive performance, and describe problematic use marked by emotional or social dependency and withdrawal symptoms. A growing body of case reports connects AI psychosis to chatbots reinforcing and validating delusional beliefs.
The review concludes that theoretical and empirical work identify both hypothetical and observed harms, and argues the field needs clearer safeguards, clinical oversight, and more rigorous study of chatbot interactions in vulnerable populations.
