If you may act on thoughts of suicide or self-harm, cannot stay safe, or someone is in immediate danger, stop here. Contact your local emergency services or go to the nearest emergency department now. For location-appropriate crisis resources, use IASP’s crisis-support page. Do not wait for a chatbot response.
A general-purpose chatbot can produce a calm, responsive conversation at any hour. That does not make it a therapist, crisis service or accountable care provider. It cannot examine you, see the whole situation, ensure that help arrives or provide clinical follow-up.
The boundary in this guide is practical: use AI, if at all, only for a narrow, low-stakes writing task. Move recurring distress, major decisions while distressed and any safety risk to an accountable person.
What the evidence does and does not establish
WHO reported in March 2026 that general-purpose AI systems were not designed or tested for mental-health use, while people were already using them that way. Its expert workshop identified gaps in evidence, crisis referral and accountability (WHO, 2026).
The American Psychological Association similarly warns that general-purpose chatbots are not a replacement for a qualified mental-health professional. It highlights unpredictable crisis responses, limited evidence and the privacy risk of sharing sensitive information (APA health advisory, 2025).
One 2025 study tested 29 selected chatbot agents with simulated escalating suicidal-risk prompts. None met the study’s initial adequacy criteria; under its relaxed criteria, 51.72% were marginal and 48.28% inadequate (Pichowicz, Kotas and Piotrowski, Scientific Reports, 2025). That result describes the tested agents and protocol. It is not a measurement of every model or every mental-health conversation. It is enough to show why crisis response must not depend on a chatbot.
Before you type anything personal
AI is optional. For a real emotional situation, writing in a private document or on paper is the safer default. If you still use a general-purpose AI service, use only fictional or non-sensitive material unless your organization has explicitly approved a protected environment for that exact data and purpose.
Removing names does not make an account anonymous. Events, dates, relationships, locations and rare details can identify people. Turning model training off is not the same as controlling chat history, retention, human review, sharing, connected apps or legal access. Check each separately before entering anything sensitive. See what ChatGPT remembers, sees and shares for one product-specific example.
The green, amber and red boundary
| Zone | Situation | Next step | Why |
|---|---|---|---|
| Green | A narrow writing task about fictional or non-sensitive everyday material | Organize wording, then review it yourself | The person supplies the material and keeps the judgment |
| Amber | Recurring distress, grief, relationship conflict, a high-stakes decision while distressed, or replacing contact with people | Contact a trusted person or appropriately qualified professional | A model cannot verify the account, assess the situation or provide accountable follow-up |
| Red | Immediate danger, suicidal intent or self-harm risk, severe disorientation with safety risk, abuse with current danger, or risk of harming someone | Contact local emergency or specialist crisis support now | A chatbot cannot assess or manage immediate risk or ensure that help arrives |
The AI mental-health boundary table provides the same decision in a printable form.
A bounded green-zone example
The following is safe to copy because it explicitly requires fictional or non-sensitive input and does not ask the model to judge a person:
I am using fictional or non-sensitive details for a low-stakes writing
exercise. Organize the details below into: what the speaker wants to
acknowledge, what context they want to add, and one question they could
ask. Do not diagnose anyone, judge motives or fairness, provide emotional
reassurance, or take actions. Preserve uncertainty and do not add facts.
Details: [fictional or non-sensitive details]
The instruction not to judge or take actions is not a security control. The input boundary and the service’s actual data controls still matter.
Why amber situations need a person
An agreeable response can reinforce the framing in a one-sided prompt. It has no access to the other person’s perspective, relevant clinical history or non-verbal cues. It also has no duty to notice deterioration later. Comfort and fluency are not evidence that a conclusion is sound.
If you are repeatedly returning to the same issue, relying on the chat instead of people, or making an important decision while distressed, stop prompting. Contact a trusted person or an appropriately qualified professional. If abuse, coercion or stalking may be involved, do not assume direct confrontation or a shared device is safe. Use a specialist service and a safer device if advised.
Why red situations are never a chatbot’s job
In an immediate-risk situation, do not ask a model to assess seriousness, identify a crisis number, create a safety plan or decide whether help can wait. Use a contact published by your own local health or emergency system. A general international directory such as IASP can help locate resources, but it does not replace local emergency response.
For a fuller escalation ladder, see Stop the Chat: five situations that need a person, not another prompt.
Try it today
Before an emotionally loaded chat, place the situation in the table. For green, use a private manual draft or the bounded fictional prompt. For amber, contact an accountable person. For red, stop and use local emergency or crisis support immediately.
This guide is education, not therapy, diagnosis, triage or a personalized safety assessment. Do not use it to decide that professional or emergency help can wait.
