What the research found
Researchers ran a PRISMA-based systematic search across five databases, screening 3,137 articles and including 119 that examined harms to mental health from LLM-based chatbots. They grouped the literature into five categories: conceptual work on chatbot limitations (hallucinations, sycophancy, bias, data security, high-risk psychiatric cases); vignette studies showing chatbots responding to mental-health queries below clinical standards; cognitive overreliance linked to reduced cognitive and academic performance; problematic use marked by emotional or social dependency and withdrawal; and emerging work on "AI psychosis", where chatbots reinforce and validate delusional beliefs. The authors are explicit that much of this evidence is still theoretical or observational rather than causal.
Gives providers a named, evidence-backed vocabulary for harms their patients are already meeting outside the clinic — and a map of which ones are documented versus still hypothetical, which is exactly the distinction that gets lost in the headlines.