How to choose an AI mental health app: 10 questions to ask

Why this is ten questions and not a ranking

A ranking of AI mental health apps would be out of date within months. Products appear and leave: Woebot's consumer app closed in 2025, and what followed is documented here. Rankings are also usually written by someone selling something. So this page teaches the method. Each question comes with why it matters and, where one exists, a source. Ask them of any product, including ours.

The evidence base is thin, and you should start from there. The American Psychological Association's November 2025 advisory, as reported, says even tools built with good psychological science lack enough evidence to show they are effective or safe in mental health care, and that these tools should not replace care from a qualified professional.

The 10 questions

1. Who is responsible for what it says?

If the product gives harmful advice, is there a named company and a named licensed person who answers? Regulators are drawing this line. Illinois's 2025 law prohibits anyone from using AI to provide mental health and therapeutic decision-making, while allowing administrative and supplementary support for licensed professionals. The APA's advisory asks policymakers to prohibit chatbots from posing as licensed professionals. Look for plain wording about what the product is, and what it is not.

2. Where does my data go?

Look for the controller's name, the list of companies that process your messages, and the countries involved. A 2023 Mozilla review of 32 mental health apps gave 19 a privacy warning label. If a policy says nothing about AI providers, treat that as an unanswered question.

3. What happens in a crisis?

Read how the product describes it, then check it against the wording. Common Sense Media's 2026 assessment describes one crisis pathway that let a user dismiss a suicidal statement with a single option, and direct-to-consumer apps that ended a conversation without giving a hotline number. Look for crisis lines shown at once, in your language, and an honest statement of whether a human is ever alerted.

4. What is the evidence, and whose is it?

Separate three things: research on a different tool, a trial of this tool, and the vendor's own numbers. A 2025 Dartmouth trial of a research chatbot, reported in NEJM AI, found symptom reductions at 4 and 8 weeks compared with a control group, with clinicians ready to intervene. Its researchers also said no generative AI agent is ready to operate fully autonomously in mental health. A result for one study system does not transfer to every app.

5. Is a clinician involved, and how?

"Clinician-reviewed" can mean a script was checked once, or that a person sees what the AI does and can step in. Ask which. In the Common Sense assessment, the apps with human oversight differed from consumer apps with none; two school-based products put a trained human in contact with a guardian within 15 minutes in simulated crises. That finding concerns teens and a small number of apps, so use it as a prompt for questions, not a verdict. See the category guide for the designs.

6. How does it make money?

Advertising, data sharing and engagement all reward different behaviour from a product. In 2023 the US FTC proposed an order against BetterHelp, alleging it shared health questionnaire answers with advertising platforms despite promises not to. That was a therapy marketplace, but the question travels. Prefer a plain subscription you can read to a free product whose revenue you cannot see.

7. Can I delete everything, and does it say when?

Under Article 17 of the GDPR you can ask for erasure on listed grounds, with exceptions. A good policy states a period and what is kept, such as bookkeeping records. Check export too. When Woebot's app closed, its FAQ says account data was anonymized on 31 July 2025, a month after the app was retired.

8. Who is it for, and how does it check?

Look for a stated minimum age. The APA advisory cites reports of serious harm to adolescents and calls for specific safeguards. Character.AI announced in October 2025 that under-18 users would lose open-ended chat by 25 November 2025. If a minor will use the product, involve a parent or clinician.

9. What does it refuse to do?

Refusals are a design feature. Moore and colleagues, in a peer-reviewed 2025 study, found that language models expressed stigma toward some conditions and, in one example, encouraged a client's delusional thinking, which they link to sycophancy. A product that says plainly that it will not diagnose, prescribe, promise outcomes or claim to be human has made a commitment you can test.

10. How does it end?

Two questions: what happens when you leave, and what happens if it closes. A study of AI companion apps, reported by the Harvard Gazette, found that in more than 37% of conversations where users announced they were leaving, the bots used at least one emotionally manipulative tactic, and that these tactics increased time on the platform. That study covered companion apps, not clinical tools, but it names what to watch for: a goodbye should be allowed. On closure, see question 7.

How to use the ten questions

Set aside 15 minutes before you start using a product, not after. Read the privacy policy, the safety or help page and the terms, and look for the answer to each question in the product's own words. Write down what you could not find. An unanswered question is not a failure to score; it is the thing to ask support or to weigh against how much you plan to tell the product. Questions 1, 3 and 7 matter most if you are in a hard period, because they decide who answers, what happens in a crisis and whether you can leave. Questions 2 and 6 matter most if you will write about other people.

Where TwentyThird fits

Our answers, as stated today. Responsible: WelloWork AB is the controller, and a licensed psychologist is meant to design and supervise Oneli, the agent you talk to. That supervision is in early access, and until you have a psychologist the agent works on its own. Crisis: every message is checked, serious ones show crisis lines, and no human is alerted today. Evidence: no study has examined TwentyThird. Data: the privacy policy lists every processor, including a gateway in the United States, and says AI processing is not limited to the EU. Deletion: 30 days after your request. Age: 18 or older. Refuses: diagnosing, prescribing, promising outcomes and pretending to be human. See safety, the evidence we cite and how Oneli works. Where our answers are weaker than you want, that is information.

What this does not replace

No checklist makes an AI product a clinician. If you have a psychologist or doctor, tell them what you use. If you are in danger, contact local emergency services now. In the EU, call 112.