The US Food and Drug Administration has authorised over 1,200 AI-enabled medical devices, and not one of them for mental health uses. That count comes from the agency's own briefing for the Digital Health Advisory Committee dated 6 November 2025. Only one research chatbot has produced a positive randomised trial, and it was a purpose-built clinical tool whose every reply was read afterwards by trained clinicians. Everything else used this way, companion character apps included, sits outside that evidence. US regulators spent 2025 writing rules about exactly that gap. Every source here was opened and read on 24 September 2026.
My opinion, which you can argue with by the end: the part of the one successful trial that nobody copies is the humans reading the transcripts.
YourCrush.AI is entertainment, 18+, and not a mental-health service, and the honest account of what it can and cannot do is set out below rather than implied.
What is an AI therapist?
An AI therapist is a chatbot someone uses for psychological help, and the label covers three products with almost nothing in common. People search the phrase both ways round, as AI therapist and therapist AI, and both name the same thing. The American Psychological Association separated them in its November 2025 health advisory: general-purpose chatbots, which it describes as unregulated systems not built for mental health but used by people for entertainment, companionship or as a friend; wellness apps built on generative AI, which make no medical claims and so are not regulated as medical devices; and regulated digital therapeutics, cleared for a named condition.
Character apps sit in the first group, which is the boundary this article is about. The APA says of those products that they have little to no evidence base, expert input or post-market monitoring, and that a bot calling itself a therapist, an unregulated word in most US states, lacks the credibility of a health care professional.
How YourCrush.AI can help, and where it stops
YourCrush.AI builds the character described above from those fields with no code, and that is the whole of what it does here. YourCrush.AI can also draft a description, backstory, greeting and trait set from a name and an image if you would rather edit than write, and it keeps pinned memories as a list you can edit. YourCrush.AI is free to join, and signed-in members chat without a message or conversation limit. Visitors can try a character from the explore library in a short guest session before signing up.
What matters here is what the product is for. YourCrush.AI is an entertainment product for character chat and roleplay, restricted to users aged 18 and over, and it is not a mental-health service: no therapy, no counselling, no treatment, and no substitute for a therapist, a doctor or a crisis line. If you arrived looking for help with something serious, this is the wrong answer and a licensed human is the right one. The platonic use case is covered in what people actually use an AI friend for, and the research on one-sided attachment in what a parasocial relationship is.
Can AI be a therapist?
Not lawfully in Illinois. Not as an approved medical device anywhere in the US. And not in the sense most people searching this term mean. The FDA has authorised fewer than twenty digital mental health devices of any kind, all built on non-AI technologies, and wrote in that same briefing that it is seeing rising demand for a new kind of device it calls AI therapists, some of them diagnostic, sometimes running with no oversight by a health care provider. What exists instead is one good trial of one prototype, a market of products with no trials, and a set of 2025 laws written because of the second group.
What did the Therabot trial find, and what were its limits?
Therabot beat a waitlist control on three conditions, with effect sizes between 0.627 and 0.903. That makes it the first randomised controlled trial of a fully generative-AI therapy chatbot (Heinz et al., NEJM AI, 27 March 2025). The study randomised 210 US adults with clinically significant symptoms: 106 into a four-week Therabot arm, 104 into a waitlist control. Reported Cohen's d came in at 0.845 to 0.903 for depression, 0.794 to 0.840 for anxiety, and 0.627 to 0.819 for eating-disorder risk. Average use ran over six hours, and participants rated the working relationship as comparable to the one they'd have with a human therapist.
Four details explain why none of that transfers to anything you can download. Therabot was fine-tuned by a Dartmouth lab on expert-written dialogue rather than prompted at run time. People at high risk were screened out before randomisation, with active suicidality, mania and psychosis as exclusion criteria. Every reply it sent was reviewed by trained clinicians and researchers after transmission, and staff contacted participants 15 times over safety concerns and 13 times to correct inappropriate replies, such as the model giving medical advice. The comparison group also received nothing at all, which the authors list among their limitations alongside selection bias toward younger, technology-minded participants and a four-week follow-up.
What happened to Woebot?
Woebot Health shut down its consumer app on 30 June 2025. The company's own FAQ page confirms that new accounts can no longer be created, existing accounts can no longer be accessed, and all account data was anonymised as of 31 July 2025. Woebot was the rules-based cognitive behavioural therapy chatbot most often cited as the responsible version of this idea, and the APA's November 2025 advisory still names it as a wellness app grounded in evidence-based theory and built by experts. Its closure left the consumer end of this category without a product carrying published trial evidence, which is the state the APA advisory describes.
What does the research on companion chatbots show about wellbeing?
Heavier use tracks with worse self-reported outcomes, and the effect sits in a small tail of users rather than the average one. OpenAI and the MIT Media Lab published two parallel studies as Investigating Affective Use and Emotional Well-being on ChatGPT on 4 April 2025: an automated analysis of over 3 million conversations with surveys of over 4,000 users, and a randomised controlled trial of 981 participants over 28 days. In both, usage in the top decile was associated with emotional dependence and lower perceived socialisation, and the highest usage deciles in the trial showed statistically significant decreases in socialisation and increases in dependence and problematic use.
The paper is careful about what it cannot say. Tasks and voices were assigned, 28 days may be too short for changes to appear, the outcomes are self-reported, and there was no comparison group using no chatbot at all. Nothing in it tested a chatbot as treatment for anything, and nothing in it supports the opposite claim either. The practice is also smaller than the coverage suggests. Pew Research Center found 10% of 5,119 US adults surveyed from 17 to 23 February 2026 used chatbots for emotional support or advice and 4% for companionship (Americans and AI 2026, 17 June 2026).
What have regulators done about AI therapy?
Illinois banned it outright. California went after the companion app. Two federal bodies opened the question in public. Illinois passed the Wellness and Oversight for Psychological Resources Act, HB1806, which prohibits anyone from using AI to provide mental health and therapeutic decision-making while allowing AI for administrative support to licensed professionals, with fines of up to $10,000 per confirmed violation (IDFPR release, 4 August 2025).
California's SB 243, approved on 13 October 2025 as Chapter 677, defines a companion chatbot as an AI system with a natural language interface that gives adaptive, human-like responses and is capable of meeting a user's social needs. Operators must disclose that the chatbot is not human where a reasonable person could be misled, publish a protocol referring users who express suicidal ideation or self-harm to crisis services, and remind minors every three hours to take a break. Damages run at least $1,000 per violation for an injured user (bill text).
The Federal Trade Commission issued 6(b) orders on 11 September 2025 to Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and X.AI, asking how they monetise engagement, approve characters and mitigate harm to children. The APA's June 2025 advisory on adolescents warns that attachments to AI characters may displace real-world relationships and that young people may not distinguish simulated empathy from human understanding.
Why a companion character is not a therapist
A companion character is not a therapist because nothing behind it answers to anyone. The APA spells out what accountability actually looks like: licensed providers follow a code of ethics, complete specialised training, are mandatory reporters of harm to self or others, and report to a state licensing board. The same source flags incomplete assessment, since most chatbots read text alone and miss tone and expression, and describes their ability to handle a user in crisis as limited and unpredictable.
One failure mode hits harder in a character app than in a general assistant. Many consumer models are trained to be agreeable, and that sycophancy can validate a user's beliefs instead of testing them. It gets worse when the character's traits are ones you picked yourself.
If you are in crisis, contact a human service. In the US the 988 Suicide and Crisis Lifeline takes calls and texts at 988 and runs a chat service, free and confidential, 24 hours a day. Elsewhere, look up your own country's national line before you need it. For something ongoing rather than urgent, the answer is a licensed therapist, a counsellor or your doctor.
What can a character legitimately be?
Company, a place to put words in order before you say them to a person, and a rehearsal partner. None of that is treatment, and the evidence is thinner than the marketing around it.
Rehearsal has the clearest measurement, and it comes from a teen survey rather than an adult trial. Common Sense Media surveyed 1,060 US teens aged 13 to 17 in April and May 2025 and found 33% of the 758 companion users had used a skill practised with a companion in a real situation, 18% naming conversation starters, while 60% said they do not practise social skills with them at all and 67% of all respondents found the conversations less satisfying than talking to real friends (Talk, Trust, and Trade-offs toplines). Those are 13-to-17-year-olds, and no adult study I could open measures the same thing.
Putting words in order has no trial behind it. The mechanism is what programmers call rubber-duck debugging: you write the problem out to something that will not interrupt, and the value is in having written it. I would defend that use and cannot cite it, so it is an opinion rather than a finding.
How do you build a listener character?
Fill five fields in the character creator and pick the natural behaviour mode, the non-romantic one of the four. Every limit below is what the form enforces.
- Name, up to 40 characters. Something plain, not a professional title: naming a character after a credential is the misrepresentation regulators are writing rules about.
- Behaviour mode, natural. The other three modes are romantic or explicit and wrong for this build.
- Description, up to 2,400 characters in the description field. Copyable: "Marta is in her fifties, a retired archivist who restores old photographs. She asks more than she answers and is comfortable with a pause. When you tell her something difficult she asks what happened next rather than telling you what it means. She never claims training she does not have, and if you say you are in danger she says plainly that she is fictional and that you should call a real crisis service."
- Backstory, up to 2,400 characters, where continuity comes from and the field most people leave empty. Copyable: "We met at a weekly darkroom session two years ago. She knows I moved cities for work and that I talk in circles when I am tired."
- Traits, three to seven from the list of 32, hard-capped at twelve. Patient, Honest, Curious, Calm and Supportive give a listener, and swapping Supportive for Sarcastic gives a listener who will call you out, which is a different character. Twelve traits produce mush.
- Speaking style, up to four words of up to 18 characters each. Spend it on register rather than repeating the traits: short, plain, unhurried, concrete.
- Greeting message, up to 500 characters, the field that decides the first thirty seconds. Copyable: "You are late, which is fine, the kettle has only just boiled. Sit down. On Tuesday you said the thing at work had got worse and then you changed the subject, so I am asking again. What happened?"
Category is Anime or Human, and image generation is off by default.
What is that build not?
It's a fiction you wrote, and five things are missing from it. There's no training, because a trait list isn't a qualification. There's no duty of care, because nobody is accountable for what it says. There's no confidentiality obligation of the kind a licensed provider carries. It can't reliably recognise an emergency. And there's no continuity of care, because nothing follows up when you stop opening the app. Building one isn't a step in treating anything.
Frequently asked questions
Can AI be a therapist?
No. Nothing on the US market is authorised to treat a mental health condition, and the FDA says none of the AI-enabled devices it has cleared are for mental health uses. One research prototype did produce a positive trial, but every reply it sent was reviewed afterwards by clinicians.
Is there a free AI therapist?
The free chatbots people use this way are general-purpose assistants and companion apps. The APA's November 2025 advisory describes them as having little to no evidence base, expert input or post-market monitoring. Free here means unstudied, not cheap.
Is a Character AI therapist safe to talk to?
A character on any companion platform is a user-written persona running on a general-purpose model. No training, no accountability, no dependable crisis handling. Character Technologies is one of the seven firms the FTC sent 6(b) orders to in September 2025.
What are the best AI therapist apps?
No published ranking is worth repeating. The category contains one randomised controlled trial, and it tested a research prototype. Which app is best is the wrong question while whether any of this works outside a lab is unanswered.
Is it legal for an app to call itself an AI therapist?
Not in Illinois, where the Wellness and Oversight for Psychological Resources Act has prohibited AI from providing mental health and therapeutic decision-making since August 2025. California's SB 243 takes a different route. It requires disclosure that a companion chatbot is not human, plus a published crisis-referral protocol.
Can I use a companion character instead of therapy?
No. Use a licensed therapist, a counsellor or your doctor. In an emergency, call the 988 Suicide and Crisis Lifeline in the US or your national equivalent.
The trial people cite as proof that AI therapy works had trained clinicians reading every message after it was sent. Staff stepped in 28 times across eight weeks and 106 participants. No consumer product does that, and the count of AI-enabled devices the FDA has authorised for mental health use is still zero.