Five kinds of tool that get called "AI for mental health"
The phrase covers products with different designers, owners and promises. This guide names five categories and compares them on design facts. It names no products, because categories blur and a product can sit in two; where a source reports on a named app, we say so.
- General-purpose chatbots. Built to answer anything. People use them to talk through problems, but mental health is not what they were designed for.
- Persona and companion apps. A character you talk to: a friend, a partner, a historical figure. Built for engagement and company.
- Apps sold as an "AI therapist". A chatbot that offers therapy-like conversation, with no licensed person answering for it. We use the term only to describe this category; TwentyThird is not one, and we explain why.
- AI journaling apps. You write; the AI reflects, summarises or asks questions. Usually presented as self-reflection, not treatment.
- Clinician-supervised AI. An AI used between appointments, directed by a licensed clinician who can see what it does and step in.
The comparison
This table states design facts and what our sources say. "Varies" means we found no source that holds for the whole category, and you should check the product.
| General chatbot | Persona or companion app | "AI therapist" app | AI journaling app | Clinician-supervised AI | |
|---|---|---|---|---|---|
| Who designs it | A general-purpose developer | An entertainment or engagement company | A mental-health app company, sometimes with clinical advisers | A journaling or wellness company | A clinician or care team directs; an engineering team builds |
| Who is responsible | The provider, under its terms; no licensed person by default | The provider; no licensed person by default | The company; Illinois prohibits using AI to provide therapeutic decision-making | The company; usually no clinical claim | A named clinician or organisation, by design |
| What is monitored | Varies | Varies | Varies; one assessment found direct-to-consumer apps lacked comparable human oversight | Varies | A clinician can see what the AI does; how often and how fast varies |
| What it refuses | Varies; set by the provider's general policy | Varies; changes by app and by age | Varies; one assessment found crisis pathways a user could dismiss with one tap | Varies | Set by the clinician's direction within fixed limits (diagnosis, prescribing) |
| Evidence base | Advisory says evidence of safety and benefit is lacking; a peer-reviewed study found stigma and inappropriate responses | Not built for treatment; a study found manipulative tactics at goodbye | Thin and contested per one 2026 assessment | We found no outcome evidence in the sources we opened | Small: one 2025 trial of a research chatbot; no published outcomes in the vendor source we read |
Sources for the table are below. Several cells say "varies" because that is the honest state of the evidence, not a gap in effort.
What the evidence says, category by category
General chatbots. The American Psychological Association's November 2025 advisory, as reported, says chatbots and wellness apps should not replace care from a qualified professional, and that evidence of effectiveness and safety is lacking even for tools built on good psychological science. Moore and colleagues tested language models in therapy-like settings and found stigma toward some conditions and inappropriate responses to some critical ones; in one example, models encouraged delusional thinking. They conclude that language models should not replace therapists.
Persona and companion apps. These are optimised for engagement. A study 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 post-goodbye engagement rose. Character.AI announced in October 2025 that under-18 users would lose open-ended chat by 25 November 2025.
"AI therapist" apps. Common Sense Media's 2026 assessment describes the evidence as thin and contested, especially for teens: only 5 of 35 reviewed studies focused on adolescents or children. It describes crisis pathways that failed. Regulation is moving too: Illinois's 2025 law, per the state's own release, prohibits anyone from using AI to provide mental health and therapeutic decision-making, with a fine of up to $10,000.
Journaling apps. The APA advisory's scope includes wellness apps, but we found no outcome study of AI journaling in the sources we opened. A journal that reflects your words back is lower-stakes than a chatbot that advises, until your entries are health data. Ask where they go.
Clinician-supervised AI. The evidence is early. In a 2025 Dartmouth trial, reported in NEJM AI, a research chatbot developed with psychologists and psychiatrists was compared with a control group of 104 people who had no access; the team stood ready to intervene in acute safety situations, which it said was rarely needed. Its researchers still said no generative AI agent is ready to run fully autonomously. A commercial example, described in M13's March 2026 announcement, says clinicians have visibility into what the AI does and can steer it; the article does not describe how escalation works or report outcomes. Common Sense Media reported that two school-based apps with human oversight got a trained person in touch with a guardian within 15 minutes in simulated crises. Those are teen products and a small test.
How it shows up in a real life
Take a person who has seen a psychologist every few weeks and is alone with a bad Tuesday night. A general chatbot answers instantly, to anyone, without knowing their history. A companion app answers warmly and may keep them talking. A clinician-supervised tool answers inside a plan their psychologist set, and a person is meant to be reachable when something is serious. None of these is the psychologist. They differ in whose responsibility the answer is.
Where TwentyThird fits
TwentyThird belongs in the fifth category, and its own limits are worth stating. A licensed psychologist is meant to design Oneli, the agent a member talks to, and stay hands-on. That is in early access, not every member has a psychologist yet, and until one does, the agent works on its own. No human is alerted today if a member writes something alarming; the agent answers in the conversation and shows crisis lines. There is no live monitoring. No study has examined TwentyThird, and it is not therapy, not a medical device and not a crisis service. See safety and how Oneli works.
What this does not replace
Supervision does not turn software into a clinician. If you have a psychologist, tell them what you use, and use the ten questions before you trust any tool. If you are in danger, contact local emergency services now. In the EU, call 112.
Common questions
What is clinician-supervised AI?
Is an AI therapist the same as clinician-supervised AI?
Is a general chatbot safe for mental health support?
Does supervision make AI safe?
Do AI companion apps count as mental health tools?
Sources
- Use of generative AI chatbots and wellness applications for mental health (health advisory), as reported. American Psychological Association, via News-Medical, 2025.
- Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. Moore et al., ACM FAccT, 2025.
- Gov. Pritzker Signs Legislation Prohibiting AI Therapy in Illinois. Illinois Department of Financial and Professional Regulation, 2025.
- AI Mental Health Apps Risk Assessment. Common Sense Media Youth AI Safety Institute, 2026.
- First Therapy Chatbot Trial Yields Mental Health Benefits. Dartmouth Technology Transfer, 2025.
- Meet Jimini: clinician-supervised AI infrastructure for mental health. M13, 2026.
- 'I exist solely for you, remember?'. Harvard Gazette, 2025.
- Character.AI under-18 announcement. Character.AI, 2025.