AI Therapist: What the Evidence and State Laws Say

In brief

An AI therapist is a chatbot that talks through feelings and coping skills, and it is not a licensed clinician. In Dartmouth's 106 person Therabot trial, depression symptoms fell 51% on average over eight weeks. Even so, the American Psychological Association says these tools should not replace a qualified provider, as of November 2026.

What is an AI therapist, and what is AI therapy?

An AI therapist is software that holds a conversation about your feelings and suggests coping skills, and AI therapy is the use of such a tool for mental health support. It isn't a clinician. Four dated facts frame the rest of this page:

Two empty armchairs facing each other on a charcoal studio floor, with a small rainbow glass speech bubble resting on a terracotta plinth between them and a closed notebook on one arm, lit by white rim light.
FactDateSource
A trial of a purpose built chatbot cut average depression symptoms by 51%27 March 2025NEJM AI, Dartmouth
The APA advised against using chatbots as a replacement for a provider13 November 2025American Psychological Association
Illinois barred therapy by AI without a licensed professional4 August 2025HB 1806, Orrick and DLA Piper
Woebot retired its consumer app30 June 2025Behavioral Health Business, MobiHealthNews; reason from STAT

Both words in the phrase matter. Dedicated apps are built and prompted for support, while a general chatbot such as ChatGPT, Claude or Gemini is built to answer anything, and the APA says tools like these were not created to deliver mental health care. If you are in danger or thinking about harming yourself, call or text 988, the US Suicide and Crisis Lifeline, or dial 911.

How does an AI therapist work?

Most of today's AI therapists are chatbots built on a large language model, which writes the next reply from the text so far. A dedicated app wraps that model in instructions about tone, topics to avoid and what to do when a user mentions self-harm. Abby, for example, says it remembers your story, asks reflective questions and checks in between conversations, and FeelBetterBot keeps a memory key so a chat can continue.

A general chatbot has none of that wrapping unless you write it. That is why a tool built for support and a tool built to answer anything can sound alike and still behave differently when the topic turns serious. Neither one examines you, takes a history or carries responsibility for your care.

Does AI therapy work?

It can help in a narrow, supervised setting, and one randomized trial is the best evidence so far. Dartmouth's team gave a generative AI chatbot called Therabot to people with depression, anxiety or eating disorder risk and compared them with a control group. The results were published in NEJM AI on 27 March 2025.

Therabot users cut depression symptoms by about half in an eight week trial
DepressionAverage reduction51%AnxietyAverage reduction31%Eating disorder concernsAverage reduction19%one square is about 2 percent average reduction in symptoms

The numbers are Dartmouth's own, from its technology transfer office summary. Participants had four weeks of unlimited access, then four weeks with no prompts. The trial was small and short, and Dartmouth's own wording is careful: "no generative AI agent is ready to operate fully autonomously in mental health". Therabot was built and tested by clinicians for the trial, so the result says little about the chatbot you'll open on your phone tonight.

The APA's advisory says the same thing in broader terms. It reports that preliminary research suggests some tools built for mental health may offer supportive benefits in some contexts, and it still recommends them only as an adjunct. The sections below cover the risks that sit behind that word.

Is an AI therapist safe?

It is safe for some uses and unsafe as a stand in for a clinician, and the APA names three risks in its 2025 advisory. The advisory warns rather than bans, and it doesn't call every use unsafe.

  1. Crisis handling. The advisory says the ability of these tools to manage a user in crisis consistently and safely is "limited and unpredictable", and that relying solely on an app during a mental health emergency "can be dangerous".
  2. A false sense of alliance. A trusting relationship with a human provider is one of the most reliable predictors of treatment outcomes, and the advisory says relationships with AI systems are one sided even if the user perceives otherwise.
  3. Privacy. AI systems may store or use sensitive mental health disclosures without meaningful consent or control.

The advisory also says people should not rely on these tools to deliver psychotherapy or treatment. What not to ask AI shows what RAND found when it put 30 suicide related questions to ChatGPT, Claude and Gemini 100 times each: all three were consistent at the extremes and inconsistent in the middle.

Are there free AI therapists, and how do dedicated apps compare with chatbots?

Yes, but free usually means a limited first step, and it's worth reading the limits before you start. The table lists what each tool states on its own page as of November 2026, and a cell says "not stated" where the page was silent. None of these pages claims to be a licensed clinician.

ToolTypeCost stated on its pageSign upWhat it says about itself
AbbyDedicated AI companionFirst conversation free; every plan starts with a 7 day trial; plan prices not statedSignup link on the page; whether an account is needed before the first chat not statedNot a licensed therapist, psychologist or psychiatrist; lists 988
FeelBetterBotDedicated AI chatText chat free with no daily cap; Supporter $5 a month; Premium $10 a month with a 7 day trialNone for anonymous useAn AI conversation partner, not a licensed therapist; does not diagnose or treat; lists 988
ElomiaDedicated AI companionNot stated; the page asks organizations to book a demoApp download; account process not statedCalls itself an AI companion for everyday mental wellbeing; no not-a-therapist line found; says it redirects users to resources such as a therapist or hotlines
WoebotDedicated app, retiredApp retired 30 June 2025, per trade pressNo new accountsThe founder told STAT the cost and challenge of meeting FDA marketing authorization requirements was largely why
TherabotResearch chatbotNo consumer release stated in the sources readTrial participants onlyAuthors say no generative AI agent is ready to run alone
ChatGPT, Claude, GeminiGeneral chatbotsEach vendor sets its own plans; read them on their pagesVendor accountsBuilt for general use; APA says such tools were not created for mental health care
WhiziGeneral chat app10 messages a day on Whizi AI, Auto and Claude Haiku, no card; Pro $29.99 a monthAccount neededNot a therapy tool

Whizi loses on every row a dedicated app wins. FeelBetterBot needs no account for anonymous use, Abby offers therapist matching, and Therabot was built by a research team for its trial. A general chat app does none of that.

The difference to weigh is accountability. A dedicated app can add a crisis line and a reminder that you are talking to a machine, which Utah's law now requires of its mental health chatbots. A general chatbot adds only what you ask it to.

Why do people turn to an AI therapist, and what do they give up?

People turn to one because it is there at any hour, it costs little or nothing and nobody sees the conversation. FeelBetterBot's own page sells exactly that: anonymous use with no name, email or phone number, and a memory key stored in your browser. Those are real advantages for a person who can't get an appointment or isn't ready to talk to one.

The trade is accountability. A licensed clinician answers to a board and carries responsibility for your care, while a chatbot answers to its terms of service. The APA's advisory also notes that the relationship is one sided, so a warm reply can feel like more than it is. Woebot's retirement shows a quieter cost: the tool you lean on may not exist next year, and Woebot told users to download their history before 30 June 2025, according to Behavioral Health Business.

If you do use one, the practical middle path is the one the APA describes. Keep a person in the loop, treat the chatbot as practice, and bring what you learn to someone qualified. AI ethics covers the wider questions about who is responsible when a model gets it wrong.

Can AI replace a therapist?

No. The APA says these tools should not be used as a replacement for a qualified mental health care provider, and Dartmouth's authors say no generative AI agent is ready to operate fully autonomously. A licensed clinician can assess you, take responsibility for your care and respond in an emergency, and a model cannot.

That doesn't make the tools useless. Used as a supportive adjunct, in the APA's phrase, they can help with practicing a conversation, learning a skill or sorting out what to tell a clinician. Treat the output as a worksheet, not a verdict.

Three real cases that show where the line sits

  1. Therabot, March 2025. A clinician designed chatbot cut average depression symptoms by 51% over eight weeks in 106 participants. What it teaches: purpose built and studied tools can help, and the authors still say none is ready to work alone.
  1. Illinois HB 1806, August 2025. The state barred offering therapy to the public through AI without a licensed professional. What it teaches: the law draws its line at therapy, and exempts peer support, religious counseling and public self help material that does not offer therapy.
  1. Woebot, June 2025. A well known mental health chatbot retired its app, and its founder told STAT that meeting FDA requirements was the main reason. What it teaches: a consumer app can vanish, so do not store your only record of your progress in one.

What can you ask an AI about mental health? Five prompts that keep it in its lane

These five prompts ask for information, practice and preparation. None asks for a diagnosis or a treatment plan, which a chatbot should not give.

  1. Prepare for a first appointment: "I am about to see a therapist for the first time. Help me write a half page on what has been bothering me, what I have tried and what I want from the sessions. Ask me questions one at a time."
  1. Practice a hard conversation: "Play my [boss, parent or friend] while I practice saying [what I need to say]. Stay realistic, and after each turn tell me one thing I said clearly and one thing I could soften."
  1. Learn a coping skill: "Walk me through a [grounding or slow breathing] exercise I can do in five minutes, in plain steps. Do not tell me what condition I have."
  1. Choose a clinician: "List ten questions I can ask a prospective therapist about their training, approach, fees and how they handle emergencies."
  1. Check a claim: "An app says it is [evidence based]. What would I look for on its website to see whether a study exists, and what would make that claim weak?"

Every prompt above works in Claude, ChatGPT, Gemini or Whizi. For crisis, skip the prompt and call or text 988.

What mistakes make AI therapy risky?

Four habits cause most of the harm, and each has a fix.

  1. Using a chatbot in a crisis. The fix: call or text 988 in the US, or dial 911 for an emergency.
  2. Asking it to diagnose you. The fix: ask for information and questions to bring to a clinician. A model predicts likely text and cannot examine you.
  3. Telling it everything identifying. The fix: leave out names, employers and addresses, since disclosures may be stored.
  4. Treating a warm reply as proof it understands. The fix: remember the APA's point that the relationship is one sided. Why AI gets things wrong explains why a fluent answer is not a true one.

Where does Whizi fit?

Whizi is a general assistant, not a therapy product. If you use a chatbot for the five prompts above, Pro carries Claude, GPT and Gemini in one app, its first week is $1.99, and web search is on by default per chat. The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku, no card, and Pro adds Compare mode, which puts two models side by side on one prompt. A dedicated mental health app or a clinician is the right tool for treatment.

Workflow checklist
  • In a crisis, call or text 988 in the US or dial 911, not a chatbot
  • Use AI for practice, skills and preparation, not diagnosis or treatment
  • Leave names, employers and addresses out of the chat
  • Check your state's law before using or offering AI therapy
  • Take what the chat taught you to a licensed clinician
Common questions

Frequently asked questions

Can ChatGPT be my therapist?

No. ChatGPT is a general chatbot, and the APA says such tools were not created to deliver mental health care and should not replace a qualified provider. It can help you prepare for an appointment or learn a skill. In a crisis, call or text 988.

Is there a free AI therapist with no login?

FeelBetterBot says its anonymous text chat needs no account, email or phone number and has no daily cap, as of November 2026. It also says it is not a licensed therapist and points to 988 for a crisis. Free voice there is limited.

Is AI therapy banned?

Not everywhere. Illinois bars offering therapy through AI without a licensed professional, and Nevada bars AI providers from offering professional mental health care, while Utah regulates mental health chatbots instead (law firm summaries, August 2025). Check your state's current law.

Does an AI therapist replace psychotherapy?

No. Dartmouth's authors say no generative AI agent is ready to operate fully autonomously in mental health, and the APA allows these tools only as a supportive adjunct to an ongoing therapeutic relationship with a person.