Is it safe to talk to AI about your problems?
A careful answer comes from the American Psychological Association. In November 2025 it published a health advisory on AI chatbots and wellness apps used for mental health. Its short answer: they cannot replace a qualified therapist, and the risks are real. It names four: a machine that agrees too easily, unreliable crisis handling, data that can end up in someone's profile, and a tool that starts to feel like a person.
The advisory does not tell people to stop. It tells them what to watch for. It also tells developers what to do about each risk, and that list works as a checklist for any app you might talk to.
In June 2026 the APA followed up with a survey of 1,242 licensed psychologists. More than three quarters said their patients now bring up AI in therapy. Nearly two in five had patients who had used it to diagnose themselves.
Below, the checklist is applied to one app. The app is Epimelea, and this page is published by it. Read it with that in mind.
What the advisory does not say
It does not say that AI therapy works. Its first recommendation is not to rely on chatbots and wellness apps to deliver psychotherapy. They should not replace a qualified provider, it says, though they may serve as an adjunct to ongoing care. Even purpose-built tools, it adds, lack large, high-quality clinical trials.
That holds for Epimelea too: no such trial exists for it. What follows is not evidence that it helps. It describes how the app is built, set against what the advisory asks.
Say it is a machine, and keep saying it
The advisory asks developers to tell users, clearly and more than once, that they are talking to an AI and not a human. It also asks that no AI present itself as a licensed professional.
Here the home page of the site says the therapist is a language model. It has no human name, no photo and no credentials. It works in one school, the existential-humanistic one, and says so.
Protect what people tell it
The advisory gives a whole section to data. People disclose more to a machine than to a person, it notes, and those disclosures are recorded and can be profiled. It asks for clear answers on what is collected, how it is stored, whether it trains models, and how to delete it.
In Epimelea the archive is encrypted on the person's device. The password never reaches the server, so the server holds nothing it can read. The archive has no account behind it: it opens with its name and a password. What a person deletes is gone.
One gap remains. The words of each turn reach the language model unencrypted, because there is no other way to answer them. The model providers keep that copy for thirty days and then delete it.
Handle a crisis on purpose
The advisory calls chatbots' handling of crises limited and unpredictable. It asks developers for clear protocols for disclosures that signal a risk of harm.
In this app two parts work independently. A separate classifier reads each message before the therapist answers. When it sees risk, the app shows a block with crisis lines the person can tap. The therapist hears the same words and responds as a therapist would. Either part can miss something, so neither replaces the other.
The app is rated 18+. It does not try to meet the advisory's separate section on children.
Reduce dependency
Here the advisory and the app part ways.
Among the design features it suggests against emotional dependency is "limiting the AI's memory to prevent the illusion of a continuous relationship". Epimelea is built on the opposite choice. The therapist remembers earlier meetings, because therapy is a continuous relationship, and a therapist who forgets you every time cannot do it.
The advisory also asks for nudges to take breaks. A meeting here has no timer, and there is no streak to keep. Nothing in the app asks a person to come back. Whether that is enough to prevent dependency is a fair question, and no study has answered it.
Audits
The advisory asks for independent third-party audits of safety, bias and data security before release. Epimelea has not had one.
The question underneath
Read as a whole, the advisory keeps returning to one worry: that a person will mistake a tool for a relationship. An app built for therapy cannot design that worry away. It can only keep it in view.
So the useful question may be less about the app than about the reader. When I talk to a machine about my life, do I know at every moment what I am talking to?
Related reading: An AI chatbot as a therapist, on why language models tend to agree with the person in front of them. The advisory raises the same problem.
Sources: APA Health Advisory on the Use of Generative AI Chatbots and Wellness Applications for Mental Health, American Psychological Association, November 2025; APA press release "Psychologists say patients are turning to chatbots as mental health professionals", June 16, 2026. Both checked 2026-09-27.
Epimelea is a therapy app built around the existential school. The therapist is a language model, the archive is encrypted on your device, and the conversation carries over from one meeting to the next.
About the app: an AI therapist for long-term therapy
This page describes how one app is built. It is not medical advice and not emergency help. If you are in danger right now, contact your local emergency number, or find a helpline in your country at findahelpline.com.