Psychiatry · Medical Literature
Specialized or General-Purpose
Abstract
This Viewpoint discusses developing a taxonomy of AI mental health harm and understanding how to assess and respond to those harms to advance AI safety.
Clinical relevance
AI chatbots are now used by millions of patients weekly for mental health support—including over 1 million ChatGPT users per week with explicit indicators of potential suicide planning—yet there remains no well-accepted taxonomy of AI-related mental health harm, making accountability and regulation nearly impossible. This JAMA Viewpoint by Nelson, Kalinich, and Torous (2026) directly addresses this gap by introducing a structured harm framework and reframing the field's central debate away from 'specialized vs. general-purpose' AI toward an evidentiary safety standard. Psychiatrists and clinicians who recommend, tolerate, or encounter patient use of AI tools must understand this taxonomy to counsel patients, document risk, and advocate for appropriate oversight.
Key findings
1. The authors introduce a two-category taxonomy of AI mental health harm: Type I harms are acute failures occurring within a single or brief exchange (e.g., missing a suicide cue, enabling disordered eating, giving inappropriate medical advice, or reinforcing delusions)—these are detectable by current benchmarks and guardrails. 2. Type II harms are insidious, cumulative harms emerging over extended, multi-turn interactions (e.g., reinforced cognitive distortions, fostered parasocial dependency, and cognitive deskilling)—these are largely invisible to single-turn safety benchmarks, which the authors argue constitutes a critical measurement gap in the field. 3. Type I and Type II harms are described as orthogonal constructs, meaning a system can score well on single-turn safety benchmarks while still generating significant Type II harm over time—rendering current safety claims potentially misleading. 4. Emerging evidence cited suggests that an external, fit-for-purpose guardrail layered atop an existing model may reduce acute (Type I) harms to nearly zero, though the cost and scalability of such guardrails remain poorly characterized. 5. The authors argue that resistance to Type II harm derives from value-anchored alignment embedded in the base model architecture—not from mental health–specific fine-tuning or empathy overlays—and on this basis express skepticism toward 'mental health–specific' AI models, drawing an analogy to the failed 'digital therapeutics' label. 6. The article calls for transparency, independent evaluation, and safety standards benchmarked against real clinical care over realistic (multi-session) timescales, rather than relying on single-turn performance metrics; the evidentiary question is reframed as: what has this system been shown to do, for whom, and over what time horizon? NOTE: As a Viewpoint, this article does not report primary trial data; no effect sizes, confidence intervals, p-values, or NNT/NNH are presented.
Practice implications
Clinicians should routinely ask patients about their AI chatbot use, assess the duration and intensity of engagement (not just content), and document this as a clinical risk factor—since Type II harms (dependency, distorted cognition, social withdrawal) accumulate longitudinally and will not be flagged by the chatbot's built-in guardrails. When evaluating AI tools for clinical recommendation or integration, practitioners should demand evidence of safety benchmarked across extended, multi-turn interactions at realistic timescales, not merely single-session benchmarks, and should be skeptical of 'mental health–specific' branding as a proxy for superior safety. Professional organizations and hospital systems should urgently engage in developing clinical standards for AI harm surveillance—modeled on pharmacovigilance frameworks—given that federal regulatory authority remains fragmented and current oversight leaves both Type I and especially Type II harms inadequately monitored.