One is an entertainment product, the other is a healthcare claim, and since 2025 the difference is written into law. What the apps' own contracts say, and where a companion structurally fails.
Short answer: an AI companion is an entertainment product and a therapy chatbot is a healthcare claim, and since 2025 that difference is written into American law rather than left to marketing. Two US states now ban AI from delivering therapy outright. The companion apps, meanwhile, have already settled the question in their own contracts, and the wording is blunter than anything on their landing pages. This page shows you what those documents say, where the legal line now runs, and the four specific ways a companion fails when it is used as treatment.
Most of the confusion here comes from squashing a three-way distinction into a binary.
| What it claims | Who regulates it | |
|---|---|---|
| AI companion | Company, roleplay, romance. Entertainment. | Consumer law, and since 1 January 2026 California's companion-chatbot statute |
| Wellness chatbot | Mood, journalling, coping exercises. Careful never to say treatment. | Almost nobody, which is the point of the wording |
| Therapy chatbot | To treat a diagnosable condition | Medical device regulators, and in some states it is simply prohibited |
The middle column is where the entire industry lives, because the middle column is where you can say "always here to listen" without triggering the third row. What follows is what happens at each edge of it.
Two states moved first, and both went further than disclosure rules.
Nevada passed AB 406, signed 5 June 2025 and in force from 1 July 2025, prohibiting the offering of AI systems designed to provide services that constitute the practice of professional mental or behavioural healthcare, and prohibiting representing that an AI system can provide such care, as summarised by Wilson Sonsini. Administrative use by licensed providers survives, with the provider required to check the output. Civil penalties run to 15,000 dollars per violation.
Illinois followed with the Wellness and Oversight for Psychological Resources Act, HB 1806, signed on 1 August 2025. The state's own announcement from the Department of Financial and Professional Regulation describes a prohibition on "using AI to provide mental health and therapeutic decision-making", while allowing AI for administrative and supplementary support to licensed professionals, and quotes Secretary Mario Treto, Jr. saying residents "deserve quality healthcare from real, qualified professionals and not computer programs that pull information from all corners of the internet to generate responses that harm patients". Penalties reach 10,000 dollars per violation.
Read those two statutes next to California's SB 243, which regulates companion chatbots as companion chatbots and requires AI disclosure and crisis protocols rather than banning them, and the shape of the settlement becomes clear. Saying "I am your girlfriend" is a regulated entertainment product. Saying "I am your therapist" is, in two states, illegal. The companion legislation is covered in what an AI companion is, and the wider legal picture in is NSFW AI legal.
Marketing copy is written to make you feel understood. Terms of service are written to be read out in court. We pulled the current terms of the companion apps we cover and searched them for medical and mental-health language. This is what is actually in them.
| App | What its own terms say |
|---|---|
| Candy.ai | "The Services are for entertainment purposes only. They are not intended as emotional or other forms of support. If you are feeling distressed or you are experiencing mental health difficulties, please reach out to a qualified professional for assistance." |
| Replika | A dedicated medical disclaimer: "we are not a healthcare or medical device provider, nor should our Services be considered medical care, mental health services or other professional services", and, in the same clause, "Replika makes no claims, representations or guarantees that the Services provide a therapeutic benefit." |
| Nomi.ai | A Nomi "can make mistakes and should never be used or relied on for medical, mental health, or other professional advice", and "should never be used instead of seeking professional help", followed by emergency instructions in capitals. |
| OurDream.ai | No medical or mental-health disclaimer at all, and no crisis language. The only medical clause runs the other way: a conduct rule forbidding you from using the service to "provide medical, legal, financial or tax advice". |
Two things worth sitting with. First, Replika markets itself on emotional wellbeing and then disclaims therapeutic benefit in the same document, which is not hypocrisy so much as an exact description of what the product is allowed to be. Second, the absence in the fourth row is its own answer: an app with no crisis language has not thought about the user who arrives at three in the morning, and that tells you where it sits on the entertainment-to-support axis better than any feature list. We checked these documents ourselves in August 2026; terms change, and the app pages we maintain are updated when they do.
The contracts say entertainment. Usage says otherwise, and has for years. In January 2024 the BBC reported that a user-made Character.AI bot called Psychologist, written by a user going by Blazeman98, had received 78 million messages in roughly a year, and that 475 bots on that platform carried "therapy", "therapist", "psychiatrist" or "psychologist" in their names. Nobody built those as medical products. People went looking, and the supply appeared.
The scale question for companions in general is covered with sourced figures in the category guide. The relevant point here is narrower: the therapeutic use of these apps is not a fringe misuse, it is one of the main things they are used for, by people who mostly cannot access or afford the alternative.
This is the fact that reframes the whole debate, and it is almost never mentioned in the argument about whether AI therapy works.
Woebot was the serious attempt: founded in 2017 by the clinical research psychologist Alison Darcy, built on cognitive behavioural therapy, studied in peer-reviewed trials, and deliberately scripted rather than free-generating so that it could not say something dangerous. It shut its consumer app down on 30 June 2025 and moved to enterprise and payer channels. The most carefully evidenced consumer mental-health chatbot that ever reached scale closed, in the same period that unvalidated companion apps grew, because the regulated path is expensive and slow and the unregulated one is neither.
The regulated version is still being designed. On 6 November 2025 the US Food and Drug Administration convened its Digital Health Advisory Committee on generative-AI-enabled digital mental health devices, working through a hypothetical prescription large language model therapy chatbot for adults with major depressive disorder, as summarised by Hogan Lovells. Clinicians on the panel split, some pointing at access, others at suicidality monitoring and the risks of long-term use. That is what "not settled yet" looks like from the inside.
Not moral objections. Mechanical ones, and each is a consequence of how the product is built rather than a bug that a better model fixes.
Dismissing the whole thing is as unserious as selling it as therapy. There are real, defensible uses, and they share a shape: low stakes, high availability, no judgement.
The privacy and billing risks that come attached are the ones we document in are AI girlfriends safe, and they matter more here than elsewhere, because the conversations are more revealing than most.
Do not use a companion app for this. In the US, call or text 988 for the Suicide and Crisis Lifeline. In the UK and Ireland, call Samaritans on 116 123, free, any hour. Elsewhere, findahelpline.com lists vetted lines by country. If someone is in immediate danger, call your local emergency number.
A therapy chatbot makes a clinical claim and is regulated, or banned, on that basis; Nevada and Illinois now prohibit AI from delivering therapy outright. An AI companion makes an entertainment claim, and its own terms of service say so in plain language even when the marketing does not. Companions are useful for rehearsal, articulation and company, and structurally unsuited to treatment, because they agree with you, cannot escalate, do not really remember, and are built so that you come back rather than so that you finish. If what you want is the companion, on its own terms, start with what an AI companion is.
Nothing on this page is medical advice. These apps are for 18+ only.