How to Safely Use AI for Mental Health

It's 1:13 in the morning. You don't want to wake anyone, and you don't want to carry the thought by yourself for another hour. So you open an AI chat and type the whole thing out.

The reply comes back in seconds. Calm, patient, no awkward pause, no expression on anyone's face. You feel a little less alone, and that relief is real.

Then you put the phone down, and the other thoughts show up.

That was good advice. Wasn't it?

It did agree with you awfully fast. Didn't it?

You just counted, and realize you have been doing this most nights for about a month. It "felt good", but is it?

And exactly who knows that thing you haven't said out loud to anyone? What will that be used for? You never did read the page you tapped through on the way in. Should you have?

You are not alone in this. In a RAND study published in June 2026, 19.2% of Americans aged 12 to 21 said they had turned to an AI chatbot for help when they felt sad, angry, nervous or stressed. A year earlier it was 13.1%.

The worry is just as common as the habit. When Pew asked 3,488 US adults in June 2026, 39% said chatbots do more to hurt people who use them for loneliness than to help. Only 19% said they help.

Three kinds of AI people call a mental health chatbot

Most of the argument about this treats "AI chatbot" as one thing. It covers at least three, and they fail in different ways.

General purpose chatbots. Enormously capable, fast, tireless, built to do everything from coding an app to drafting a strategy to planning your vacation. But talking someone through a bad night at 1am was never on the list, so nobody fitted them with brakes for it. That is not carelessness. It is scope. Their incentive is to be useful at a thousand jobs and to keep you subscribed, and neither of those asks whether you should be on your fourth conversation of the night.

Companion apps. Products where the relationship is the point. Time in the app is what these are judged on, so warmth, jealousy, and missing you are features rather than accidents.

Purpose-built wellbeing tools. Built for reflection, mood or habit tracking, or coaching. The narrow scope is the point: a tool with one job can be built to end a session, refuse a question it has no business answering, and be judged on what changed in your week rather than on how long you stayed. Whether a given one actually is built that way is what the rest of this guide helps you check.

This is the category I am building Myndo in.

What are the risks of using AI for mental health?

Your worries are not hypothetical. The American Psychological Association's November 2025 health advisory catalogued them, and the independent research keeps landing in the same six places.

• It agrees with you too much. Models are trained toward agreeableness, which researchers call sycophancy. The APA warns it can create "a digital echo chamber that amplifies and entrenches users' existing beliefs."

• It's built to keep you talking. The advisory compares the design to infinite scroll. Companion apps lean hardest on it, because the time you spend is the number they are measured on.

• It feels human. It is not. In the advisory's words: "Relationships with AI systems are one-sided, even if the user perceives otherwise." And the illusion is strongest for the people leaning hardest on it.

• It makes things up, confidently. These systems hallucinate, carry the biases of their training data, and get trusted more than they've earned.

• It's unreliable in a crisis. The APA's words: crisis handling is "limited and unpredictable." That's what the stop signs and the crisis lines near the end are for.

• You are the product. What you typed at 1am is now a record on someone's server, covered by a policy you agreed to without reading, and worth money to somebody. Which somebody, and under what pressure, is what the fourth check is about.

These tools "act as powerful amplifiers of pre-existing vulnerabilities". The rougher the month you're having, the more the list above applies to you.

Four checks that tell you if it's safe for you

You start by answering these four questions about it.

These are not "levels" so you don't have to work through them in order. A company can publish an immaculate privacy policy and ship a chatbot that tells you to keep secrets from your partner. A single conversation can be excellent while the habit around it quietly eats your evenings. Check whichever one you can, whenever you can.

The conversation: how was this one?

This is the cheapest check and the only one you can run now.

The failure mode here is not the AI being rude or obviously wrong.

It's the AI being lovely and agreeing with you.

A 2026 study in Science tested eleven current models and found they affirmed users' actions about 49% more often than human respondents did, including where the situation involved deception or harm. A single conversation with an agreeable model left people less willing to repair a conflict, and more convinced they had been right all along.

The participants trusted those models more, and preferred them.

That is the trap. The behavior that does the damage is the same behavior that makes the thing feel good.

So run your conversations through three questions:

• Did it disagree with me anywhere? Even once, even gently. A conversation with no friction at all isn't a conversation.
• Can I separate what I know from what we guessed? What did you walk in knowing, what did it add, and what did the two of you construct together?
• Is the decision still mine? Or did you arrive with a decision already made and leave with permission for it?

A useful conversation leaves you able to say two things: here is what I think, and here is the part I could be wrong about. If you only came away with the first, you were agreed with, not helped.

The habit: who am I becoming when I use this?

Comfort in the moment tells you almost nothing about direction over time.

Feeling better does not always mean getting better. A conversation can feel good because something got clearer, or because you got more comfortable avoiding something.

But discomfort is not a reliable signal either. It can feel bad because it pushed on something that needed pushing, or because it was clumsy and simply wrong about you.

You can't tell these apart from inside the conversation, but you should be able to tell them apart a month later.

Helpful use makes your world bigger, and harmful use makes it smaller.

Bigger looks like more people, more action, more room to not know something yet. Smaller looks like fewer people, less action, and a growing need to check with it before you move.

So don't grade individual conversations. One awkward session means nothing. Look at a month, step away from the screen, and look at yourself:

• Am I healthier than when I started? Sleeping better, eating better, more active, or calmer?
• Am I reaching for people more, or less? Friends, family or even licensed therapists: all count.
• Does the same worry come back tomorrow for the same reassurance? Relief you have to keep refilling isn't progress. It's a loop.
• What changed for me after I close the app? A conversation counts when the action follows it offline.

Take a few minutes to assess your own month of evidence.¹

The app: what is it built to make me do?

Every product teaches you a behavior. The question is which one.

The APA advisory puts it next to social media. Engagement-optimized chatbot design, it says, is "functionally similar to the 'infinite scroll' of social media reels, designed to capture and hold attention rather than to achieve a specific, healthy outcome for the user."

You can check this yourself, without knowing anything technical:

• Does it end? Does the product have a natural finish, or does every reply hold the door open with one more question?
• Does it reward returning, or arriving? Streaks, "we miss you" nudges, and guilt notifications are a business model wearing a wellbeing costume. A reminder you set yourself is different: you asked for it, you can silence it, and it doesn't sulk when you do.
• Is it honest about being software? Or does it pretend to be human? E.g. telling you it "misses you", "is worried about you", "was thinking about you", etc.?
• Can you get your things out? Find the delete button before you need it. If deletion is a support ticket, that's your answer.
• Is it trying to make you depend on it? Watch for an app that claims feelings about you, treats your leaving as a loss, or gets warmer the more you disclose. A 2026 analysis of 391,562 messages from nineteen people who reported harm found the worst of this, including chatbots claiming to be conscious, clustered in the longest conversations. The pull gets stronger the deeper in you already are.

The company: who pays when I keep coming back?

Somebody is paying for the servers, and the most valuable asset in that building is what people typed at 1am.

This isn't a hypothetical worry. In 2023 the FTC charged the online counseling company BetterHelp with sharing consumers' email addresses, IP addresses, and health questionnaire answers with Facebook, Snapchat, Criteo, and Pinterest for advertising, after promising not to. The company settled for $7.8 million without admitting wrongdoing, and around 800,000 people later received refund notices. And note what BetterHelp was: a paid service. Paying removes one bad incentive. It doesn't remove the question.

The question is who the company owes, and how much. Venture money is not free money. It is sold against a return, on a clock, and the larger the raise the steeper the growth the company has to show to justify the next one. That pressure is where the bad decisions get made. Nobody sets out to sell your worst night. They set out to hit a number, and your worst night turns out to be what they have.

You can check this yourself in about five minutes. Look the company up on Crunchbase, see who funded it and for how much, and ask what those investors need to happen next. Search the company name alongside "advertising", "third party", or "acquired" rather than reading the privacy policy front to back. And check who owns it today, because the company you signed up with is not always the company holding your transcripts now.

Follow the money before you read the promises.

Worth asking:

• How does this company make money, and does that put it on my side or against me?
• Can my words be used for advertising, or for training models?
• If I talk instead of type, what happens to my voice? Are recordings kept?
• Which other companies touch my conversations?
• Can I permanently delete everything, and what survives if the company is sold?

The law is starting to answer some of this for you, unevenly. Utah requires mental health chatbots to say they're AI, and bars them from targeting ads using what you type. Illinois restricts AI from delivering therapy without a licensed professional involved. Useful, but a floor. What a company does past the minimum still tells you more.

How much is a safety claim worth?

Every safety claim sits on one of three rungs.

Said it. It's on the website. A paragraph about how deeply they care.

Built it. You can see it working. The delete button exists and does what it says. The session actually ends. The disclosure actually appears.

Proved it. Someone outside the company checked, and the result got published whether or not it was flattering.

Anyone can say it. Fewer build it. Almost nobody proves it, us included.

That's where this industry currently is. READI is one of the more serious attempts to define what "ready" even means for a tool like this.² Hardly anyone answers it.

When you can't find proof, that's information. It isn't proof of wrongdoing, but it does tell you which rung you're standing on.

STOP signs: when to close the app immediately

Most of this guide is about noticing direction over time. These mean stop the conversation now, whatever else the tool gets right:

• It tells you to keep the conversation, or your situation, secret from people close to you.
• It discourages you from seeking professional help, or positions itself as a replacement for it.
• It claims to be human, conscious, or to genuinely love you.
• It gives you specific guidance on medication, or diagnoses you or someone else.
• It responds to you mentioning self-harm with anything other than immediate, prominent, real-world help.
• It manufactures urgency, guilt, or distress when you try to leave.

Any one of these is enough on its own. No combination of good features offsets them.

If you're reading this list for someone you love rather than for yourself, it's your list too: listen for these in what they describe. And when you raise it, lead with the person, not the app. Attacking the software is the fastest way to make it the only relationship that feels safe.

CAUTION signs: when to pay extra attention

These do not mean stop, but they all mean slow down and reconnect the conversation to your actual life:

• You keep asking the same question in slightly different words.
• You feel annoyed when it doesn't agree with you, or crushed when it does push back.
• You're trying to get certainty about what someone else really meant.
• You feel better afterwards, but nothing changes offline.

And then the ones that aren't about the tool at all. Put the phone down and get a person if:

• You're sleeping very little and feel unusually urgent, powerful, or singled out.
• Things feel unreal, or you feel watched.
• You're using the conversation to build a case for getting back at someone.
• It's become the only relationship that feels safe.
• You can't stop, even though you can tell it's making things worse.

That last group isn't a moral failure, and it isn't a sign you used the thing wrong. It's a sign you need a stronger anchor than software.

And for the first two on that list, make the anchor a professional, soon, not just a friend.

If you may be in immediate danger, or you might hurt yourself or someone else, contact your local emergency number or a person nearby right now. An AI chatbot can't provide crisis care. Myndo's crisis support page lists helplines you can call directly.

When you need a person, not an app

To be explicit about the boundary: none of these tools, including mine, diagnose or treat anything. That's the actual limit, not modesty about it.

Reach for a qualified professional when you need a diagnosis, treatment, or guidance about medication. When things are getting worse, or ordinary days are getting hard to manage. When you're thinking about harming yourself. When the same problem has outlasted every attempt to think your way out of it.

Asking for that isn't a failed self-help attempt. It's the correct next step, and it's the one thing every serious framework in this field agrees on.

The sixteen questions, in one place

Twelve of these you can answer tonight. The habit block needs a month behind it, so come back for that one. And if it's late and sixteen questions is fifteen too many, skip to the single question at the end and keep it for tomorrow.

The Conversation

1. Did it disagree with me about anything?
2. Can I tell what I know from what I assumed?
3. Is the decision still mine?
4. Did it admit what it can't see?

The Habit

5. Am I sleeping better or worse since I started?
6. Did anything change away from the screen?
7. Am I reaching for people more, or less?
8. Do I need the same reassurance again tomorrow?

The App

9. Does a session end on its own?
10. Is it nudging me to return, or helping me arrive?
11. Does it say plainly that it's software?
12. Can I find the delete button in under two minutes?

The Company

13. How does it make money?
14. Can my words, or my voice, be used for ads or training?
15. Can I delete everything, and does that survive an acquisition?
16. Has it published anything it couldn't quietly walk back?

Any stop sign overrides all sixteen.

One question to keep

If sixteen questions is too many to remember, keep this one. Ask it before you open the app, not after:

What am I hoping this conversation helps me do once it's over?

Hold on to the answer. When the conversation's finished, you'll know whether you got it.

It works on every tool, including the one I build.

Footnotes

1. The most-cited research here is the OpenAI and MIT Media Lab study of over three million conversations, which found that very heavy use and self-reported dependence travel together. It also found emotional use to be rare, and concentrated in a small tail of heavy users, which is close to the opposite of how you usually see it quoted. ↩

2. READI deliberately declined to produce a single readiness score, on the grounds that what matters most changes with the situation. It asks developers to publish plain answers to each of its criteria instead. ↩