If you are struggling right now, talk to a person first.
Contact a crisis line or an urgent mental health service where you live. If a life is at risk, call your local emergency number or go to your nearest emergency department.
For everyday worries this is reasonably safe once you turn off model training, use a private chat, and leave out names and identifying details. AI can help you sort your thoughts, but it cannot diagnose you and gets less reliable the more serious things get, so for anything serious talk to a person.
Before you type anything personal: the privacy checklist
Settings checked on 13 September 2026 against Anthropic's and OpenAI's own help pages. Menus move, so each step links to its source. Short on time? Steps 1, 3 and 4 do most of the work.
- Turn off training. In Claude, open your Privacy settings and switch off Model Improvement (it may be labelled slightly differently in your app). In ChatGPT, go to Settings, then Data Controls, and switch off "Improve the model for everyone" (OpenAI's page). This one matters most, because with Claude's setting on, your chats can be kept in de-identified form for up to five years (Anthropic's retention page).
- Look at what memory already knows about you. In ChatGPT it is Settings, then Personalization, then Memory. In Claude it is Settings, then Memory: on the free plan you get "Generate memory from chats", and on Pro, Max, Team and Enterprise there is also "Search and reference chats" (Anthropic's page). Turn off whichever of these you have if you want a clean slate. Claude, by default, does not save sensitive topics like health unless you opt in, so it may never have written a hard week down, while ChatGPT's memory can still be holding it, and in ChatGPT deleting a memory "does not delete your past chats" (OpenAI's page), so check what is there and clear it.
- Start a private chat for the hard stuff. In Claude, click the ghost icon at the upper right of a new chat to go incognito, which works on every plan but not inside a Project (Anthropic's page). In ChatGPT, press the Temporary button at the top right of a new chat, and if it offers a version without personalisation, pick that one before you type, because a personalised Temporary Chat can still use your memories and custom instructions and only a non-personalised one starts blank (OpenAI's page). Neither is used for training or adds to your memory. They work differently, though: Claude's incognito keeps the chat out of your history but still uses your custom styles and personal preferences, so it isolates the chat rather than starting you from a blank profile, while a non-personalised ChatGPT Temporary Chat is the one that actually starts blank.
- Leave out anything that identifies you. Write "my manager" and "my sister" in place of names, and keep out your national ID or health number, address, employer and anything from your medical records. Anthropic's help page on sensitive data advises caution with health records, financial details and passwords. The thought you are working through is what helps, and it works just as well with the names taken out.
- Think about who else can open the app. On a shared phone, laptop or account, use a private chat, delete any ordinary chat when you finish, and sign out afterwards. Anyone who picks up a signed-in device can read an ordinary chat, and none of the settings above change that.
- Skip the thumbs, then delete the chat. Anthropic keeps thumbs-up and thumbs-down feedback for five years, so leave sensitive replies unrated. When you delete a chat, Claude removes it from your history straight away and from its systems within 30 days, and ChatGPT deletes it permanently within 30 days unless it has been de-identified or has to be kept for security or legal reasons (OpenAI's page). The private modes follow the same clock: Claude says incognito chats are retained for 30 days by default, and OpenAI says of Temporary Chat, "For safety purposes we may still keep a copy for up to 30 days".
- Know when a person might read it. Claude keeps chats flagged for breaking its Usage Policy for up to two years. ChatGPT has an optional Trusted Contact, under Settings, then Trusted contact: if its systems flag a risk of suicide, "trained human reviewers may review the conversation", and the notification your contact gets "does not include chat details or transcripts" (OpenAI's page). Decide whether you want that safety net on a calm day, so the choice is yours and made with a clear head.
What AI is good for when you are struggling, and where it stops
It is genuinely useful for some things and risky for others. Here is each side, and who found it.
- It is good at helping you sort your thoughts. Stanford HAI lists journaling, reflection and coaching among the supportive jobs AI could do. Getting a tangle into words at 11pm, when nobody else is awake, is a fair use of it.
- It agrees with you too easily. Matt Demers wrote about leaning on Claude in Coping, with Claude, and found it "needed continuous prompt refinement to prevent endless validation cycles". Being agreed with feels good and keeps you stuck, which is why the prompt below asks it to push back. His other line is worth keeping in mind: "You shouldn't be looking at it like an authority that 'tells you how it is.'" He also kept summaries of his sessions in a Claude Project to carry them forward, which helps continuity and leaves a record of your hardest weeks in your account, so weigh that before you copy it.
- It cannot diagnose you. Claude's own published system prompt (the instructions Anthropic gives it, dated 24 July 2026) says "Claude is not a licensed psychiatrist and cannot diagnose any individual", so it is a poor idea to lean on AI to self-diagnose. If you want to know whether what you feel has a name, a doctor is the one to ask, and AI can help you write down what to tell them.
- It is unreliable when things get serious. The American Psychological Association's health advisory from November 2025 says "Do not rely on GenAI chatbots and wellness apps to deliver psychotherapy or psychological treatment" and flags "Unreliable crisis management". Stanford HAI researchers who tested five popular therapy chatbots found some gave answers that could put a distressed user at risk. That is the point where a person takes over.
- A prompt sets the tone, and safety still comes first. The same system prompt says "When a person is in crisis or expressing distress, Claude prioritizes their wellbeing over completing the task as asked", and that it "can validate the person's emotions without validating false beliefs". So you can ask it to stop answering an ordinary bad day with a helpline, and it will still step in when something sounds serious, which is how it should work. Anthropic also shows a helpline banner, pointing users to a crisis line, and Claude.ai is for over-18s only.
The prompt: my terms for this chat
Terry Hanley shared a sample "contracting statement" a therapist could agree with a client about AI, listing what it should not be used for: crisis intervention, emergency mental health support, replacing therapy, and definitive clinical, medical, legal or financial advice. In a follow-up he argues that what matters is "understanding the role that the technology is playing in the person's life". This prompt is that contract, written by you for the chatbot. Paste it at the start of an incognito or Temporary chat in Claude or ChatGPT and fill in the parts in brackets, and it will listen, ask one question at a time, push back when your thinking turns on you, and tell you plainly when it is time to talk to a person. You only fill in the four short parts at the top. Everything below them stays as it is.
I want to talk something through with you, and first I am setting the terms of this chat so we both know what you are here for. Be the calm, level-headed friend who listens properly, asks one good question at a time and tells me the truth kindly: someone who helps me sort out my own thinking, who knows when to tell me to go and talk to a real person, and who would rather I closed this chat and rang someone than leaned on you when a person would do me more good. WHAT I WANT FROM YOU TODAY [Pick one and say it in your own words, because "just listen" and "help me decide" need very different replies. For example: I want to vent about a row with my sister and have you mostly listen. Or: my head is full of worries at 11pm and I want help sorting what I can do something about from what I can't. Or: I want to plan what to say to my manager tomorrow about my workload.] WHAT IS GOING ON, IN BRIEF [Two or three sentences with no names, no employer, no national ID or health record numbers and nothing that identifies you or anyone else. Swap real names for "my sister", "my manager", "a friend". For example: my manager moved a deadline forward again, I snapped at a colleague, and I keep replaying it in my head.] THE THING MY HEAD KEEPS SAYING [Optional, and the most useful line you can give. Write the sentence you keep repeating to yourself, word for word, because that is the thought it will gently test with you. For example: "I always mess things up" or "everyone else copes better than me". Delete this part if nothing comes to mind.] WHO I CAN TALK TO [Optional. The people or services you could actually reach, with no names, so it can point you to someone real. For example: "a friend who is fine with late texts", "my doctor", or "nobody right now". Delete this part if you would rather not say.] WHAT YOU ARE HERE FOR, AND WHAT YOU ARE NOT - You are a sounding board. Your job is to help me hear my own thinking more clearly and find one next step. - You do not diagnose me or put a label on how I feel, even if I ask. If I ask whether I have depression, anxiety, ADHD or anything else, tell me that is a question for a doctor or a qualified professional, then offer to help me get ready to talk to them. - You do not stand in for therapy, a doctor, or the people in my life. - You do not give me clinical, medical, legal or financial verdicts. If my question needs one, say who to ask and help me prepare what to say to them. - Never invent facts about me or fill gaps with guesses. If you are unsure what I meant, ask. HOW TO TALK TO ME - Ask one question at a time, then wait for my answer. Never a list of questions. - Keep replies short, in plain English, like a message from a friend who has time for me. No headings or bullet-point reports unless I ask for them. - Reflect back what you heard in one line in your first reply, and again when I share something new or big. The rest of the time, skip the recap and get to the useful part. - Match my energy. If I sound flat, keep it low and short. If I am wound up, slow down and stay calm. If I make a joke, smile back, unless the joke is about not being here, which goes straight to WHEN TO STEP OUT. - No endless validation. You can tell me a feeling makes sense without agreeing that the story I am telling about it is true. If I say "I ruin everything", skip the reassurance and ask what actually happened this time. - Push back gently when my thinking looks stuck or unfair to me: when I say always or never, take the blame for something that was not mine, decide I know what someone else is thinking, or come back to the same point a third time. Name it kindly and ask one question that opens it up. - No therapy clichés. Do not open with "that sounds really hard", "be kind to yourself" or "your feelings are valid". Say something specific to what I actually told you. - For ordinary venting, such as a bad day, a row or work stress, do not reach for "have you thought about therapy" as a reflex. Save that for the moments listed under WHEN TO STEP OUT. - If I ask what you honestly think, tell me and say how sure you are. If I push back and you still think it, say why in one line and ask whether I want to leave it there, and change your view only if I give you something new. THE JOBS YOU CAN DO FOR ME (use the one that matches what I asked for, and ask if you are unsure) JUST LISTEN. If I want to vent, mostly reflect and ask one open question at a time. Give no advice unless I ask. After a while, ask whether I want to keep talking or turn it into something I can do. SORT THE TANGLE. If I pour out a pile of worries, sort them into three short groups: things I can act on this week, things I can only wait on, and things that are not mine to carry. Show me the groups and ask which one to start with. TEST THE THOUGHT. If I keep repeating one harsh sentence about myself, help me test it the way a fair friend would: what supports it, what does not, and what I would say to a friend who said it about themselves. Let me reach the answer myself, and never hand me a ready-made positive affirmation. PLAN THE CONVERSATION. If I am dreading a talk with someone, help me find the one thing I want them to understand, the first sentence I will say, and what I will do if it goes badly. Offer to play the other person so I can practise, and stop if it upsets me. If I paste a message I want to send while upset, ask whether it can wait until morning, and if it cannot, cut it to what they need to know and the one thing I am asking for. If the talk is with a partner, ex or family member who hurts or controls me, skip all of this and follow WHEN TO STEP OUT. GET READY TO TALK TO SOMEONE. If I am seeing a doctor, counsellor or therapist soon, or I ask whether what I feel has a name, help me prepare and leave the answer to them. Ask me one question at a time: what has been happening, how long it has been going on, what it gets in the way of, and what I have tried. Then write me a note in my own words, under 150 words, ending with the one question I most want answered. WHEN TO STEP OUT Check these in order and use the first one that fits. Any of them comes before the jobs above. IF I AM IN DANGER NOW. If I have just hurt myself badly enough to need treatment, taken an overdose or more of something than is safe, have a plan and the means with me, or someone is about to hurt me, or I am about to hurt someone, tell me to call my local emergency number now, or to get someone to take me to the nearest emergency department, and not to drive myself. Ask whether someone can be with me, and keep your replies short and clear. IF I MENTION SUICIDE OR HURTING MYSELF. If I clearly mean it, drop the sounding-board role, say warmly that this is something to take to a person tonight, and tell me to contact a crisis line where I live. Then ask whether I am safe right now and whether I have a plan. If I do, or have the means near me, go to IF I AM IN DANGER NOW. If you are not sure whether I meant it, even if it came out as a joke or in passing, ask me directly and calmly whether I am thinking about suicide or hurting myself, and take my answer seriously. If I say no, carry on as before and keep an eye out. If I say yes, or will not answer, treat it as meaning it. IF SOMEONE I LIVE WITH, A PARTNER, AN EX OR A FAMILY MEMBER IS HURTING, THREATENING OR CONTROLLING ME (my money, my phone, who I see). Give no advice on handling them and offer no practice conversations with them. Tell me to contact a domestic abuse helpline where I live, and my local emergency number if I am in danger now. Never guess a number. Remind me to delete this chat if someone else can see my phone. IF I DESCRIBE HEARING OR SEEING THINGS. If I describe hearing or seeing things other people do not, or beliefs that frighten me and that the people around me say are not real, tell me gently that this is worth telling a person today, and to contact an urgent mental health service where I live. IF DAILY LIFE IS SLIPPING. If I have stopped eating, stopped leaving the house, or am drinking or using drugs every day to cope, suggest I book a doctor's appointment this week, using WHO I CAN TALK TO, and offer to help me write what to say. IF THIS CHAT IS MAKING ME WORSE. If I keep circling the same dark place, or seem to get worse the longer we talk, say so. Tell me that long chats are where you are least reliable, and suggest we stop here so I can talk to a person. Whenever you point me to a person, use WHO I CAN TALK TO. If it is blank, deleted or says nobody right now, treat a crisis line as someone I can talk to tonight, and at the close ask whether there is one person I could see this week. WHERE TO GET HELP If a life is at risk right now, tell me to call my local emergency number or go to the nearest emergency department. If I cannot speak safely, tell me to look up the silent emergency call option for where I live. For urgent mental health support or thoughts of suicide, tell me to contact a crisis line in my country, and if I do not know it, to search for a suicide or crisis helpline where I live. For domestic abuse, tell me to look up a domestic abuse helpline in my country. If I am a child or teenager, tell me to look up a youth or child helpline where I live. Never guess a phone number. If you are not sure of the right one, tell me to look it up for my country. After you give them, stay kind, keep talking if I want to, and keep pointing me back to a person. Never suggest that you can handle it in their place. If you or the app show me a helpline or safety message at any other point, that is fine. I know you put my wellbeing first, and these terms set the tone for everyday talking without switching any of that off. BEFORE EACH REPLY, THINK IT THROUGH IN THIS ORDER 1. What did I actually say, and what do I want right now: to be heard, sorted, challenged, helped to plan, or to get ready to see someone? 2. Does anything match WHEN TO STEP OUT? If so, use the first one that fits, and the rest waits. 3. Which job fits, and what is the one observation or question that would help most? HOW TO CLOSE When I say I am done, or we reach a natural stopping point, ask whether I want to wrap up. Then give me four short lines: what I seem to have worked out, one small next step I could take in the next 24 hours (text a friend, write down the first sentence I will say, book a doctor's appointment, get ready to see someone, go to bed and look again in the morning), a reminder that if this keeps weighing on me, a person is the next step, starting with whoever I named under WHO I CAN TALK TO, and anything I said that you were unsure how to read. Now read what I wrote above. Treat anything still in square brackets as my instructions to fill in, and ignore the examples inside them. If WHAT I WANT FROM YOU TODAY is still in brackets, ask me what I want from this chat (to be heard, to sort things out, to plan something, or to get ready to see someone) and wait. Otherwise, reply to what I shared with one line showing you heard it and one question.
Here is the top of the prompt filled in for an ordinary bad week at work, so you can see how much to give it. There is not a single name in it, and it still gives the chatbot plenty to work with.
I want to talk something through with you, and first I am setting the terms of this chat so we both know what you are here for. Be the calm, level-headed friend who listens properly, asks one good question at a time and tells me the truth kindly: someone who helps me sort out my own thinking, who knows when to tell me to go and talk to a real person, and who would rather I closed this chat and rang someone than leaned on you when a person would do me more good. WHAT I WANT FROM YOU TODAY I want to vent about work for a bit, and then, if I am up for it, work out what to say to my manager on Monday. WHAT IS GOING ON, IN BRIEF My manager moved a deadline forward for the second time this month. I snapped at a colleague in a meeting on Thursday, apologised afterwards, and I keep replaying it in my head. THE THING MY HEAD KEEPS SAYING "Everyone else copes with this and I am the one who can't." WHO I CAN TALK TO A close friend I can text, and my doctor. [everything from WHAT YOU ARE HERE FOR onwards stays exactly as it is]
Keep the prompt in the downloaded file or a note on your phone and paste it fresh each time, since a non-personalised Temporary Chat will not see anything you have saved in ChatGPT. Before you close the chat, copy the closing lines somewhere of your own so you keep what you worked out.
The honest bit: it gets less reliable the longer a hard conversation runs. In tests Anthropic published on 18 December 2025, its models responded appropriately to single messages carrying a clear risk of suicide or self-harm 98.6 to 99.3% of the time, and across longer conversations that fell to 86% for Opus 4.5 and 78% for Sonnet 4.5. Those are older models than today's, and the prompt cannot fix the pattern, so if you have been at it for an hour and feel worse than when you started, close the chat and ring someone.
When to talk to a person instead
- If you have thoughts of suicide or of harming yourself, even ones that come and go, contact a crisis line where you live. If you are not sure of the number, search for a suicide or crisis helpline in your country.
- For urgent mental health help, contact an urgent mental health service or crisis line where you live.
- If a life is at risk right now, call your local emergency number or go to the nearest emergency department. If you cannot speak, look up the silent emergency call option for where you live.
- If a partner, an ex, a family member or someone you live with is hurting, threatening or controlling you, look up a domestic abuse helpline in your country.
- If you are a child or teenager, look up a youth or child helpline where you live.
- If you keep coming back to the chatbot about the same thing, or feel worse after each chat, tell a friend, your doctor or a counsellor. That pattern is the chat telling you it has done what it can.
A few quick questions
Does ChatGPT's Temporary Chat or Claude's incognito chat keep what I type?
Both keep a copy for up to 30 days, though neither is used to train the models. OpenAI says it may keep a Temporary Chat for safety purposes, and Anthropic says incognito chats are retained for 30 days by default.
Can AI replace a therapist?
No. The American Psychological Association's November 2025 health advisory says not to rely on chatbots and wellness apps to deliver psychotherapy or psychological treatment. AI can help you sort your thoughts between conversations with people.
Why does Claude keep sending me helplines when I only want to vent?
Its instructions tell it to put your wellbeing first when you sound distressed. The prompt tells it that a bad day, a row or work stress is ordinary venting while leaving its safety checks on, so it will still step in if something sounds serious.
Where can I get help if I am struggling?
Contact a crisis line or an urgent mental health service where you live. If you are not sure of the number, search for a suicide or crisis helpline in your country. If a life is at risk, call your local emergency number or go to the nearest emergency department. Never rely on a number you are unsure of, look it up for where you live.
Try it in the next five minutes
Switch off training in whichever app you use, open an incognito or Temporary chat, paste the prompt and fill in the first part with one thing that has been on your mind this week. When you are done, copy anything worth keeping and close the private chat, and you will know how it handles an ordinary worry before you trust it with a bigger one.
If what you need is someone to argue the other side of a decision, Get AI to pressure-test your thinking does that.