Is It Safe to Talk to AI About Mental Health? Where the 2 a.m. Confession Goes

Do any of these sound familiar?

• It's 1:40 a.m. and you're typing a chatbot a paragraph you haven't managed to say to your best friend in three years.
• You ask an AI to help you untangle a fight with your partner and paste in the whole thing: names, quotes, and the bit where you said something you regret.
• You close the tab feeling like that stayed between the two of you. You, and a server in another state.

Then, a day later, a quieter question turns up: where did all of that actually go?

Why it's so easy to tell a chatbot everything

A chatbot doesn't sigh. It doesn't check the clock, raise an eyebrow, or bring up what you told it six months ago at a family dinner. There's no face to watch for disappointment and no risk that it'll think less of you. It's like talking to a stranger on a train, the kind you'll never see again, except this train writes down every word.

That's why the chat window feels private. For most of us, privacy is a social feeling: we judge it by who's in the room, and nobody visibly is. Whether a conversation is actually private depends on what the company stores, who can get to it, and what the law says, and none of that shows up in the text box.

Fun fact: the person running the best-known chatbot has said this out loud. In July 2025, on Theo Von's podcast This Past Weekend, OpenAI's CEO Sam Altman pointed out that when you talk to a therapist, a lawyer or a doctor, "there's legal privilege for it," and added, "we haven't figured that out yet for when you talk to ChatGPT," as TechCrunch reported. A CEO's remark carries no legal weight, but it's a frank one.

What actually happens to what you type:

Every company does this differently, so take ChatGPT as one example, with setting names from OpenAI's help center at the time of writing.

• It stays in your history until you delete it. Your conversations sit in your account, ready to scroll back through, by you or by whoever picks up your unlocked phone.
• It may help train future models unless you switch that off. In ChatGPT the setting is called "Improve the model for everyone," under Settings, then Data Controls, according to OpenAI's Data Controls FAQ. Turn it off and new conversations stop being used for training.
• Temporary chats play by different rules. According to OpenAI's Temporary Chat FAQ, they don't appear in your history and aren't used for training, though a copy may be kept for up to 30 days for safety reasons.

Deleted isn't always deleted

In May 2025, as part of The New York Times' copyright lawsuit against OpenAI, a court ordered the company to preserve output logs that would otherwise have been deleted, including chats users had deleted themselves. According to OpenAI's own statement, that covered ChatGPT Free, Plus, Pro and Team, and also data that companies sent through OpenAI's API without a zero data retention agreement. So apps built on top of OpenAI were in it too, unless they had that agreement. The obligation ended in September 2025, with exceptions. You don't need to panic over it, but it's a good illustration that "delete" depends on the law as well as on the button.

If you're weighing up a mental health app rather than a general chatbot, the ten minute privacy policy check walks through what to look for in its policy.

What research turned up:

In 2014, Gale Lucas, Jonathan Gratch and colleagues at the University of Southern California published a study in Computers in Human Behavior built on health-screening interviews with a virtual human, an animated interviewer on a screen. Some participants were told it was fully automated. Others were told a person was operating it behind the scenes. The people who believed they were talking to a computer reported less fear of disclosing and worried less about the impression they were making, and they were more willing to open up.

The thing that makes it easier to talk is the same thing that makes you forget who's listening. The limit: this was a virtual interviewer in a lab, years before ChatGPT existed.

In February 2024, the Mozilla Foundation reviewed 11 "romantic" AI chatbots for its Privacy Not Included guide. Every single one got the warning label, and ten of the 11 failed Mozilla's minimum security standards, such as requiring strong passwords. It's a nonprofit's review of product policies rather than a peer-reviewed study, and these were apps built for intimate talk.

In November 2025, the American Psychological Association published a health advisory on generative AI chatbots and wellness apps used for mental health. It said these tools don't yet have enough evidence to show they're effective or safe for mental health care, that they shouldn't replace care from a qualified professional, and it called for "safe-by-default" settings and comprehensive data privacy legislation. Your therapist's notes have a law behind them. Your chat history has terms of service.

The rules are moving too: in September 2025 the Federal Trade Commission asked seven companies, OpenAI among them, how their companion chatbots affect children and teens. Check the settings, not this article.

Is it safe to talk to AI about mental health? Five things to do first:

1. Find the training switch before the first heavy conversation. In ChatGPT it's under Settings, then Data Controls. In other tools, search the settings for "training" or "improve the model." Make it a rule: "Settings first, feelings second."
2. Use the temporary mode for the hardest stuff. That's the mode for conversations you don't want to reread in a year, and don't want anyone else rereading either.
3. Leave out names, places and anything a stranger could match to a real person. "My manager" works as well as a full name and job title, and initials will do for everyone else. The AI doesn't need your ex's surname to help you think.
4. Decide what goes where. Numbers leave a small footprint: a 4 for sleep, a 7 for anxiety, a 3 for energy. Whole paragraphs leave a big one. And not every hard evening has to be typed out at all. Sometimes the first step is calming your body down without typing a word.
5. Keep your therapist for the things that need a person. A therapist is bound by professional confidentiality and can hear what you're not saying. An AI can help you sort your thoughts before a session. It can't do that part.

Wellbeing check-ins
Three numbers instead of three paragraphs give you the trend with almost nothing written down, which is the smallest footprint a mood record can have. Try it free in MindSync →

Since you're reading this on an app's blog, here's how MindSync answers the same questions, according to MindSync's privacy policy (last updated July 25, 2026). To create transcripts and insights, your entries go to OpenAI in the United States through its business API, and under MindSync's agreement with OpenAI your data isn't used to train its models. The audio from voice notes and session recordings is deleted once the transcript exists. There are no ads in MindSync, and your personal data is never sold. The same goes for MindSync: what you write in the app is protected by its privacy policy and by privacy law, not by therapist-patient privilege, so read that policy as critically as you'd read anyone else's. Delete Account is on your Account screen in the app, and the policy says your personal data is permanently erased within 30 days of deleting your account, backups included. There's more on how your entries are stored in the FAQ.

Privacy isn't a feeling you get from a chat window. It's a setting.

This is what MindSync's Wellbeing check-ins are for. Log your sleep, your anxiety and your energy as three numbers instead of three paragraphs. Give it time, and Mood & focus-area trends turns those ratings into something you can see: the shape of your month, without a single sentence about who said what. Then decide which part actually needs words, and where you want those words to live.

What's the last thing you typed to an AI that you'd be uncomfortable seeing printed on paper?

Deciding what to write, and where

• Reflect (AI journaling) - A guided reflection that runs through OpenAI's business API, under an agreement that keeps your entries out of its model training.
• Between-session voice notes - Speech becomes text you read and accept before it is saved, so you see exactly what was kept.
• Mood & focus-area trends - Your check-in ratings drawn as a line across the weeks, so the pattern shows up without a word written about it.

Frequently asked questions

Is it safe to talk to ChatGPT about mental health?

It can be, if you know what's kept, who can read it and what you've switched off. Check the "Improve the model for everyone" setting, use Temporary Chat for the most personal topics, and remember that ChatGPT conversations don't carry the legal privilege a therapist's do. Settings differ between AI companies, and a chatbot isn't a replacement for a therapist.

Are conversations with AI chatbots confidential?

Not in the legal sense. Conversations with a therapist, doctor or lawyer carry legal privilege, and in July 2025 OpenAI's CEO Sam Altman said that hasn't been worked out for ChatGPT. What happens to your chats depends on each company's policy and settings, so read those before sharing anything you'd want protected.

Does ChatGPT use my conversations for training?

It can, unless you turn it off. According to OpenAI's help center, the setting is "Improve the model for everyone" under Settings, then Data Controls, and switching it off stops new conversations from being used for training. Temporary chats aren't used for training at all. Setting names change, so check the current version of the app.

Can AI chat logs be used in court?

They can be requested in legal proceedings, because chats with an AI don't have the privilege that protects conversations with a therapist or lawyer. In 2025, a court ordered OpenAI to preserve ChatGPT logs, including deleted ones, during The New York Times' lawsuit; that obligation ended in September 2025, with exceptions. This is not legal advice.

How can I use AI for journaling more privately?

To use AI for journaling more privately, turn off model training, keep the hardest topics in a temporary or no-history mode, and leave out names and identifying details. Before you start, read the tool's privacy policy to see who processes your entries and whether they're used for training. Logging moods as ratings instead of paragraphs also leaves a much smaller record.