Center for Practical AI

CPAI Issue Brief · Health & safety

AI Use & Mental Health

Millions turn to AI in distress, and consumer chatbots carry no professional duty of care — the gap between a wellness product and a clinician.

What’s happening

People increasingly bring emotional and mental-health needs to general-purpose AI and to purpose-built companion apps. Some of these interactions reach people in crisis — and the tools they reach are not clinicians and are not regulated as such.

What the evidence shows

OpenAI has disclosed that roughly 0.15% of ChatGPT's very large weekly user base shows explicit indicators of suicidal planning or intent in a given week — over a million people (a company-reported, classifier-based estimate, attributed to OpenAI's own reporting). The American Psychological Association has advised that treating AI as a therapist or companion "carries real risk and is not advised."

The clinical evidence is narrow: therapy-chatbot studies show small-to-moderate short-term effects that fade by a few months, mostly for structured, purpose-built tools with clinician oversight — not consumer chatbots. Research on companion AI links intensive, self-disclosing, substitutive use to lower wellbeing, while also finding null effects at the average — a genuinely mixed picture.

Where it reaches constituents

Teens and adults in distress, and their families. Surveys find a majority of teens have used AI companions, and a meaningful share have turned to AI over people for serious conversations — often without knowing what the product is, or isn't, obligated to do.

The current legal & regulatory landscape

Character.AI litigation produced a May 2025 ruling treating a chatbot as a "product" for liability purposes, with settlements reported in January 2026 (terms undisclosed; no adjudicated causation). California's SB 243 (effective January 2026) sets disclosure, crisis-protocol, and minor-protection requirements for companion chatbots; Illinois restricts AI-delivered therapy; the FTC has opened a 6(b) study of companion chatbots (a study, not enforcement).

Considerations policymakers are weighing

  • ·Disclosure requirements — whether a user is told they are not talking to a licensed professional.
  • ·Crisis-protocol standards for what a product must do when a user signals danger.
  • ·Age verification and minors' access.
  • ·What clinical-validation evidence should be required before a product may claim to provide "therapy."

Listed as live debates, not recommendations. CPAI does not take a position on how these should be resolved.

This brief condenses a full, sourced public guide. The complete evidence and citations:

AI Chatbots and Mental Health →

Key sources

Survey / correlationalZhang, Zhao et al. (2025) — The Rise of AI Companions: Human-Chatbot Relationships and Wellbeing ↗Survey of 1,131 adult Character.AI users plus 4,664 donated chat sessions. Companionship-oriented use was consistently associated with lower wellbeing — strongest when use was intensive, highly self-disclosing, and offline social networks were small. Many users did not seek companionship but drifted into it. Cross-sectional, single platform.
Survey / correlationalNakagomi et al. (2026); Guingrich et al. (2025); Latikka et al. (2026) — Companion-chatbot wellbeing: moderators and null results ↗Three studies that complicate simple harm narratives: a 14,721-adult survey found companion use associated with higher wellbeing overall, peaking at moderate real-world social engagement; a 21-day randomized trial (n=183) found no overall effect on social health; and a six-country survey (n=5,663) found distress predicts use, with direction unestablished. The wellbeing moderators are loneliness and social ties — whether AI adds to or replaces them.
Randomized trial · preprint, not yet peer-reviewedKirk et al. (2025) — Dose-Dependent Impacts of Human-AI Relationships ↗Four-week longitudinal randomized trial (N≈3,534) that directly manipulated how relationship-seeking a model was. Hedonic appeal of relationship-seeking AI declined over four weeks while attachment markers and intentions to seek future AI companionship grew — a liking/wanting decoupling, with moderate relationship-seeking maximizing both. Preprint.
Randomized controlled trialFang, Liu, Pataranutaporn et al. (2025) — Investigating Affective Use and Emotional Well-Being on ChatGPT ↗MIT Media Lab and OpenAI, preregistered RCT, N=981. Harms tracked voluntary heavy use and emotional dependence — not assigned use.

A nonpartisan resource

The Center for Practical AI is a nonpartisan 501(c)(3) nonprofit. We provide education, research, and analysis, and we offer briefings and testimony on request. We do not endorse candidates or lobby for or against specific legislation. Everything here describes the evidence and the current landscape — the policy choices are yours.

Want CPAI to brief your office on this?

We provide nonpartisan briefings, research summaries, and testimony on request.

Request a briefing →