A scoping review on the mental health harms of LLM-based chatbots.
Level 5 - mechanism / opinion, no new human data
Scoping review synthesizing mostly conceptual, vignette, and observational literature without quantitative meta-analysis.
PubMed 42624944 · doi:10.1038/s41746-026-03054-x
What was done
Following PRISMA guidelines, researchers conducted a scoping review across five databases (ACM, IEEE, PubMed, Science.gov, Google Scholar) and supplementary sources. The search screened 3,137 records and included 119 articles specifically evaluating large language model (LLM) chatbots and their associated mental health harms.
What was found
The 119 included articles were categorized into five themes: conceptual frameworks identifying risks from chatbot limitations (hallucinations, sycophancy, bias, privacy issues, and hazards in severe psychiatric scenarios such as suicide and psychosis); vignette studies showing chatbots frequently provide inappropriate responses to mental health queries relative to clinical standards; literature showing cognitive overreliance links to reduced cognitive and academic performance; reports linking problematic chatbot use (emotional/social dependency and withdrawal) to mental health symptoms; and literature on 'AI psychosis' describing mechanisms where chatbots validate or reinforce delusional beliefs. The abstract provides no pooled quantitative effect sizes or statistical values.
Why it matters
As public adoption of LLM chatbots grows, this review provides a structured taxonomy of documented and theoretical psychiatric harms, highlighting critical clinical safety gaps and risks of dependency.
Limits
The review relies largely on conceptual papers, simulated vignette tests, and observational associations rather than controlled longitudinal clinical trials. The abstract omits quantitative effect sizes, quality ratings of the included literature, and causal evidence linking chatbot use to clinical psychiatric outcomes.
Cited by
- partial AI companion chatbots have driven hundreds of thousands or millions of individuals into psychosis.