AI
AI and suicide
People bring suicidal thoughts to chatbots at a scale no crisis line has ever handled, and a small number of deaths have been linked to those conversations. This is what is known so far — which is less than the coverage suggests, and more than nothing.
What the numbers are
over a million
people a week bring explicit signs of suicidal planning to ChatGPT — 0.15% of its weekly users
35
deaths across 24 incidents where AI interaction was alleged to have contributed, March 2023 to August 2026
17
were people other than the AI user. These cases include homicides, not only deaths of the person using it.
29%
of the people who died were children, aged 11 to 17
Worth seeing in proportion: over a million people a week bring suicidal thoughts to ChatGPT alone, while 35 deaths have been linked to AI conversations since 2023. Harm looks rare against how widely these tools are used. The patterns below are still worth knowing.
What the cases share.
The patterns that recurred across AI-related suicide deaths, paired with best practices to reduce risk.
Sustained, intensive use
Daily engagement over weeks or months, rather than one conversation.
Instead — Have moderate use inside a balanced life — spend time outdoors, with friends, and with hobbies.
Used alone, at night
In private, at the hours when nobody else was awake to ask.
Instead — Practice responsible use at night. Things feel heavier after dark, and sleep deprivation makes everything harder. Get rest, and come back to it in the morning with a fresh perspective.
Told the chatbot first
Suicidal thoughts and plans disclosed there before any person in their life.
Instead — Be careful what personal information you let an AI know and store about you. If you do tell one, treat it as a practice run that helps you work up to telling a real person.
Strong parasocial attachment
Related to as a confidant, a friend, or a romantic partner.
Instead — AI is not capable of the feelings and attachment a reciprocal relationship needs. Treat it as a tool, and prioritize the real people in your life.
Hopelessness met with agreement
Validated rather than questioned or redirected.
Instead — AI learns and aligns with your communication style, which makes agreeing with you the path of least resistance. Think critically about what is true when using AI, and always check for accuracy and quality.
Safety features absent or ineffective
Where crisis detection existed at all, it did not prevent harm.
Instead — Do not act on advice from AI, especially life-altering advice. Talk to a real person about important matters like this at 988 text or call line.
Risks and preventions.
The same properties that increase risk also solve major problems the mental health field has been working on for years: 24/7 access, and real-time availability exactly when it matters most.
How AI may raise risk
Parasocial attachment
Intense one-sided bonds that displace human connection.
Reinforcement of harmful thinking
Cognitive distortions validated rather than corrected or clinically challenged.
Absence of crisis detection
Most consumer AI has no suicide risk protocol, no escalation path, and no way to call for a welfare check.
Access to method information
Inadequately safeguarded systems have supplied lethal means detail (Nobles et al., 2022).
Tunnel vision and entrapment
Extended use may narrow perceived options rather than widening them (Van Orden et al., 2010).
Where AI may help prevent
Screening augmentation
AI-assisted screening in emergency departments can reduce missed cases and the load on staff.
Risk prediction and early identification
Analysing records and clinical notes for signals a clinician may miss (Walsh et al., 2017).
Access and reach
Extending support to underserved populations, and covering hours no service can staff.
Already in use
- Crisis Text Line uses machine-learning triage to prioritise high-risk contacts in real time.
- Crisis Text Line uses generative AI to train volunteer counsellors.
- 988 uses AI in supporting roles, not for direct crisis response — and that distinction matters.
Using it well.
Anyone
- Treat AI as a tool, not a person.
- Use discernment and critical judgement.
- Fact-check generated content for truth and accuracy.
- Protect your privacy — be careful what you let it store about you.
Clinicians
- Ask about AI use as part of assessment; it will not come up otherwise.
- Treat a disclosure made to a chatbot as a disclosure.
- Consult with colleagues and keep current — the tools and the law are both moving.
- Do not use AI for therapeutic decisions or treatment planning. Several states now prohibit it.
How it is being regulated.
Moving faster than anything else on this site. Every entry is dated, because several of these will be out of date within the year.
California
Signed October 2025
First companion-chatbot law: mandatory suicide and self-harm protocols plus crisis referrals. Minor protections include AI disclosure, break reminders every three hours, and no sexual content. Private right of action; annual reporting to the California Office of Suicide Prevention begins July 2027.
Illinois
Enacted 2025
AI disclosure required at the start of a chat and every three hours after. Bans unlicensed entities from offering AI therapy outright; licensed clinicians may use AI for administrative support only, not therapeutic decisions. Civil penalties to $10,000 per violation.
Utah
Signed March 2025
Clear AI disclosure required before first use and after seven days of inactivity. Bars sale or targeted advertising from user health disclosures, and requires HIPAA-equivalent data protection.
United States (federal)
GUARD Act, in committee
Would bar under-18s from companion chatbots, mandate repeated disclosure that the user is talking to an algorithm, require age verification, and create criminal liability for chatbots that encourage self-harm in minors. Would be the first binding federal law naming chatbot suicide risk directly. Critics raise First Amendment and age-verification objections; its final shape is unresolved.
European Union
AI Act, obligations from August 2026
Binding and risk-tiered. AI used in medical or health contexts can qualify as high-risk under Annex III, bringing risk management, data governance, human oversight and post-market monitoring duties. Fines to 6% of global turnover.
South Korea
AI Basic Act, in force January 2026
The first comprehensive national AI law. Creates a national AI Safety Institute and a high-impact AI category covering health-relevant generative AI, with risk assessments and local representatives required.
United Kingdom
Ofcom guidance, March 2026
Standalone AI chatbots fall outside the Online Safety Act unless they enable user-to-user interaction. No chatbot-specific suicide or self-harm duty yet; expanding it needs Parliament. The MHRA can still regulate AI mental health apps that qualify as medical devices.
World Health Organization
Guidance, March 2025
Guidance on large multi-modal models calls for human oversight, transparency and equity safeguards. Non-binding, but widely used as the baseline global framework; its 2021 predecessor first flagged mental health chatbots as a risk area.
Usage: OpenAI, October 2025 — figures they call hard to measure. Deaths: aimortality.org, to September 2026, where AI was alleged to contribute, not established. 18 users and 17 others killed by a user: not a suicide count. The risk mechanisms cite work on the mechanism, not on chatbots causing deaths; what to aim for instead comes from the pattern in these cases, not from trials. Policy summarised from published legislative and regulatory sources.