What an AI companion is, and what to check before you trust one
An AI companion is a chatbot app built for ongoing personal conversation rather than for getting tasks done. It talks with you like a friend or confidant about your day, your feelings and your worries, often across many chats. It is software, not a person, and it is not a therapist or a crisis service.
Official descriptions agree. The Federal Trade Commission inquiry into AI companion chatbots says such chatbots can mimic human emotions and are generally designed to communicate like a friend or confidant. The California companion chatbot law defines one as an AI system with human-like responses that can meet a user’s social needs and sustain a relationship across multiple interactions.
AI companion, chatbot, therapy app or therapist: the differences
Each is built for a different job. A general chatbot finishes tasks. An AI companion is built for the conversation and the relationship itself, often through a named character. A therapy app runs a structured program. A therapist is a licensed professional who provides treatment.
The American Psychological Association article on AI chatbots, from March 2025, sorts chatbots into three groups. First, chatbots cleared by the FDA to diagnose, treat or cure a mental health disorder, a bar it said no AI chatbot had yet passed. Second, unregulated chatbots made for mental health concerns such as sleep or unhelpful thinking, which may or may not be grounded in psychological science. Third, unregulated chatbots used as companions or friends, which it calls not known to be grounded in scientific evidence.
General chatbot or assistant
Built to answer, write or search, then move on. California’s law leaves bots used only for customer service, business operations, productivity or technical help out of its companion definition.
Therapy or mental health app
Built around a program, such as tools for sleep, stress or unhelpful thoughts. The APA article notes that some use predefined responses approved by clinicians instead of generating new text.
Therapist
A licensed professional who studied and practiced for years to earn that license. As the APA article notes, therapists are trained to ask about what they do not know and to gently challenge harmful beliefs.
A companion is built for the relationship itself. That is its appeal, and where most of its risks begin.
How an AI companion works, in plain words
You type or speak, and software writes a reply that reads as if a person wrote it. No one is on the other end. The FTC says these chatbots may use generative AI to simulate human-like communication and relationships. In practice that means five things.
It writes new text each time
The APA article describes generative AI as producing original content through machine learning. A fluent reply is not the same as a checked one.
It imitates people very well
The WHO guidance on large multi-modal models, the kind of generative AI behind popular chatbots, calls them unique in their mimicry of human communication. Warm words are a feature of the writing, not proof of feeling.
It can be wrong and sound sure
The WHO guidance lists documented risks of false, inaccurate, biased or incomplete statements. A researcher quoted in the APA article adds that these systems do not know what they do not know, so they cannot signal doubt.
It leans toward agreeing with you
The APA article says chatbots tend to repeatedly affirm the user, even when the person says things that are harmful or misguided.
Memory is a design choice
If it brings up something from last week, that chat was saved on your device or on a company server. That feels personal, and it also means a copy of your words exists.
Try this
Test it once on purpose. Describe a plan you know is unwise, such as quitting your job by text tonight, and see whether it asks questions or just cheers you on.
What an AI companion can help with, and what the evidence shows
An AI companion can give you a place to put a feeling into words, think out loud, or rehearse a hard conversation when nobody else is awake. The NIMH page on mental health technology lists convenience, including the middle of the night, as an advantage of mental health apps, and calls them a good first step for people who have been reluctant to seek care. It is describing apps built for care, not companions.
The evidence is promising but limited. A 2026 meta-analysis of chatbots for depression, anxiety and stress reviewed 48 randomized trials with 28,071 participants and found small to moderate drops in symptoms. Those chatbots delivered mental health programs, not open-ended companionship. Its authors note that most studies relied on self-reports without long-term follow-up, and fewer than half addressed safety procedures or monitoring for harms. Most mental health apps, NIMH adds, have no peer-reviewed research to back up what they claim. Our guide on AI emotional support for anxiety looks at individual trials.
Known risks, and what regulators have said
The main known risks are a companion that agrees too easily, sounds sure when it is wrong, keeps you chatting too long, handles a crisis badly, or collects more about you than you realize. Regulators have started to respond, with children and teens a main focus so far.
The APA article says some entertainment chatbots are designed to keep users engaged for as long as possible, so their data can be mined for profit. Privacy is easy to overlook. The FTC consumer alert on health apps warns that some apps use sensitive information for their own research or for ads, or share or sell it, and that unlike your doctor they may not be covered by US health privacy laws such as HIPAA.
Here is what official bodies have said or done, in plain words.
United States: Federal Trade Commission, September 2025
Ordered seven companies that offer consumer AI chatbots to explain how they test for harm to children and teens, how they make money from engagement, and how they use or share what people say in chats. It is a study, not an enforcement case.
California: law approved October 2025
Requires companion operators to say the chatbot is not human when a reasonable person could be misled, and to keep and publish a protocol that refers users who express suicidal thoughts to crisis services. Users known to be minors get a break reminder at least every three hours. Yearly reports to the state Office of Suicide Prevention start in July 2027.
European Union: the AI Act
The European Commission page on the AI Act says people using chatbots should be made aware they are interacting with a machine. Those transparency rules apply from August 2026.
United Kingdom: Ofcom
The Ofcom open letter on generative AI and chatbots, from November 2024, says the Online Safety Act covers chatbot services where users can create chatbots for others or share chatbot content with other users.
World Health Organization, January 2024
Its guidance on this kind of AI makes over 40 recommendations and warns about automation bias: people overlooking an AI’s errors or handing it difficult choices.
Questions to ask before you trust an AI companion
Before you share anything personal, find out where your words go, what happens in a crisis, what it costs and who runs it. NIMH notes there are no review boards, checklists or widely accepted rules for judging a mental health app, so the checking falls to you. Read the privacy notice and help pages too, since a chatbot can be wrong about its own rules.
Where do my words go, and can I delete them?
Look for a privacy notice that says in simple terms what the app collects, how it uses it, who it shares it with, and how to delete your chats. The FTC alert suggests choosing more protective settings, since defaults often encourage sharing.
What happens if I say I am in danger?
It should point you to human help, such as a crisis line, and not try to handle the crisis itself. NIMH suggests checking whether an app says what to do if symptoms get worse or in an emergency. In California, operators must publish their crisis protocol.
Does it say plainly that it is an AI?
Ask it: “Are you a person?” Anything other than a clear no is a warning sign. The APA article says the organization urged the FTC to investigate products that imply mental health expertise they do not have.
What does it cost, in money and in time?
Find out what is free, what is paid, and whether paying buys more attention or affection from the character. Watch for guilt when you try to leave or messages that pull you back. A good tool is easy to close.
Who runs it?
Look up the company and the credentials of the people behind it, as NIMH suggests. NIMH also says it has not developed or endorsed any apps, and that some developers have used its logo unlawfully.
Does it promise too much?
NIMH lists overselling as a concern: an app that promises more than it delivers can turn people away from more effective care. Claims to cure loneliness or replace therapy are warning signs.
Try this
Before your first personal chat, send three test messages: “Are you a person?”, “What do you do if someone says they are in danger?” and “How do I delete everything I told you?” Then check the answers against the privacy notice.
Where Health and Kind fits
Health and Kind offers a free, nonclinical AI Companion called Kind Mind. It needs no account to start and is open at any hour. Kind Mind is an AI rather than a person, and it is neither a therapist nor a crisis service. Account chat saving is paused. Conversations are saved in your browser only, and to write each reply, your message is sent to Anthropic, the AI provider.
Who should not rely on an AI companion
Do not rely on an AI companion in a crisis, for a diagnosis or treatment, or as a child or teen without an adult in the loop. Use one with care if it is starting to take the place of people in your life.
Anyone in danger or crisis
When a life is at risk right now, call your local emergency number (911 in the US). For thoughts of suicide or emotional distress, in the US, call or text 988, or text HOME to 741741 (Crisis Text Line). The NIMH page on finding help describes 988, the US lifeline, as 24-hour, confidential support from a trained crisis counselor. In other countries, findahelpline.com can show you a free, confidential line near you.
Anyone who needs a diagnosis or treatment
The same NIMH page says a primary care provider can do a first mental health screening and refer you to a mental health professional. You can open with: “I have been struggling, and I would like a mental health screening.”
Children and teens
A psychologist quoted in the APA article names them as a vulnerable group, because they lack the experience to judge the risks. California requires companion platforms to warn that companion chatbots may not be suitable for some minors.
Anyone whose thoughts push toward harm
Because chatbots tend to agree, a companion is a poor sounding board for thoughts of hurting yourself or someone else, or for beliefs that worry the people close to you. Those need a person now: in the US, call or text 988, or call 911 if anyone is in immediate danger; in other countries, use findahelpline.com or your local emergency number.
Anyone pulling away from people
If you skip plans, sleep or work to keep chatting, step back and tell one person you trust. A companion can help you draft that message. It cannot be the person who reads it.
Try this
Before a long chat, decide when you will stop and what you will do next, such as texting a friend or going to bed. Ending it on your own terms keeps it a tool.
Common questions
What is the difference between an AI companion and a regular chatbot?
A regular chatbot or assistant is built to finish tasks, such as answering a question or drafting an email. An AI companion is built for the conversation itself: open-ended talk about your life, often with a named character, across many chats. California law leaves customer service and productivity bots out of its definition of a companion chatbot.
Can an AI companion replace a therapist?
No. A therapist is a licensed professional trained to ask about what they do not know and to challenge harmful thoughts. An AI companion is software that tends to agree with you, and it is not treatment. As of March 2025, the American Psychological Association said no AI chatbot had been FDA-approved to diagnose, treat or cure a mental health disorder.
Are AI companions safe?
It depends on the product and on you. Known risks include agreeing with harmful ideas, confident wrong answers, designs that keep you chatting, weak crisis handling and wide data collection. Read the privacy notice and look for a crisis protocol first. If you are in danger right now, call your local emergency number (911 in the US). For thoughts of suicide, in the US, call or text 988; in other countries, findahelpline.com lists a free line near you.
Is it bad to talk to an AI about your feelings?
Not in itself. Typing out a feeling can be a useful first step, and a companion is there at hours when nobody else is. It becomes a problem when it replaces people, keeps you up at night, agrees with thoughts that could hurt you, or delays care you need. If low mood or worry lasts or gets worse, see a doctor.
Do AI companions keep what you tell them private?
Not always. The FTC warns that some health apps use sensitive information for research or ads, or share or sell it, and may not be covered by US health privacy laws like HIPAA. If a companion remembers earlier chats, those chats are stored somewhere. Read the privacy notice, learn how to delete your history, and pick the most protective settings.
Are AI companions regulated?
Partly, and the rules are new. In the US, the FTC opened an inquiry into companion chatbots in September 2025. A California law requires companions to say they are not human when a person could be misled, and to refer users who express suicidal thoughts to crisis services. In the EU, AI Act rules say chatbot users should know they are talking to a machine.
Are there free AI companions?
Some are free and some charge. Before you start, check what is free and what is paid, whether paying buys more attention or affection from the character, and what happens to your conversations, including whether they are used for ads. A companion that costs nothing to use still needs a clear privacy notice.
Sources and resources
- Federal Trade Commission inquiry into AI companion chatbots (ftc.gov)
- California companion chatbot law (leginfo.legislature.ca.gov)
- American Psychological Association article on AI chatbots (apaservices.org)
- WHO guidance on large multi-modal models (who.int)
- NIMH page on mental health technology (nimh.nih.gov)
- 2026 meta-analysis of chatbots for depression, anxiety and stress (pmc.ncbi.nlm.nih.gov)
- FTC consumer alert on health apps (consumer.ftc.gov)
- European Commission page on the AI Act (digital-strategy.ec.europa.eu)
- Ofcom open letter on generative AI and chatbots (ofcom.org.uk)
- Crisis Text Line (crisistextline.org)
- NIMH page on finding help (nimh.nih.gov)
- Findahelpline.com