A lot of people wind the day down through a chat app.
Someone's on the sofa, typing out how the day went, and the reply that comes back feels attentive, immediate, like someone's actually listening.
Increasingly, nobody's listening. Not a friend, a partner, or a family member. It's an AI companion app on the other end, built to hold exactly that kind of ongoing, relationship-style conversation.
These apps are not niche. As a category, AI virtual companions had more than 100 million registered users and more than 500 million downloads worldwide as of 2025 (Frontiers in Psychology, 2025). Replika markets itself directly as a friend, romantic partner, or mentor, and it's one of the best-known names doing it.
Character.AI has passed 500 million downloads on its own, letting people create and chat with AI "characters" (OpenReview study).
Most of the millions of conversations happening on these apps daily are entirely unremarkable: company after a long shift, someone to talk a problem through with, a low-stakes chat before bed.
Teenagers use these apps heavily, and that, combined with how popular they've become generally, is why several US states have now passed laws aimed specifically at companion chatbots. New York, California, and Washington each decided the category needed its own rules, targeting what regulators saw as the two biggest gaps: whether the app makes clear it isn't human, and what happens once a conversation turns serious.
This guide covers what actually counts as an AI companion app, why lawmakers stepped in, and what New York, California, and Washington now require. The practical part, what it means for someone who uses one of these apps or knows someone who does, comes at the end.
1. What counts as an AI companion app
How a companion app is different from ChatGPT or Claude
ChatGPT, Claude, and Gemini are general-purpose assistants, built to help with a task and then get closed. Someone drafts an email, summarises a document, or asks for a recipe, then doesn't open the app again until the next thing comes up. There's no real relationship to speak of. It resets every time.
Flip that around and there's a companion app: relationship persisting and deepening over time, not resetting.
That's the actual design brief, and the marketing too, sustain the relationship across sessions rather than complete a task and move on. Earlier conversations get remembered. The persona stays consistent. From the outset, it's framed as a friend, partner, or confidant, never really as a tool.
That design difference carries straight into the law. New York's version only applies to systems "designed to sustain a relationship across multiple interactions", and it explicitly excludes anything used solely for customer service, internal operations, employee productivity, or technical assistance (NY State Senate, S3008).
Washington's definition boils down to roughly the same test as New York's: an app has to hold adaptive, human-like, relationship-sustaining conversations to count. Ordinary business tools, gaming bots, virtual assistants, and narrow educational tools stay outside it (Washington State Legislature, HB 2225 summary).
Forget "does it use AI." The real question is whether it's built to keep a relationship going.
The apps this category actually includes
The best-known names in this category:
- Replika, positioned as a friend, romantic partner, or mentor
- Character.AI, for creating and talking to custom AI "characters", including fictional personas
- Pi, from Inflection AI
- Kindroid and Nomi, both following a broadly similar model
Woebot and Wysa sit outside this list. They're built around structured, CBT-based mental health support, not an ongoing personal or romantic relationship.
Companion apps aren't therapy tools. They shouldn't be judged against clinical evidence standards, even though plenty of people going through a hard time do use them.
2. Why several states stepped in with new laws
The safety concerns regulators pointed to
Teenagers make up a big share of who's actually using these apps. Roughly a third of teens use an AI companion app, according to research, and Character.AI users specifically averaged 93 minutes a day on it in 2024, well past what's normally recommended for screen time (OpenReview study).
That level of sustained, emotionally engaged use raised two connected concerns for lawmakers.
First: some apps weren't making it clear enough, often enough, that the user was talking to software and not a person.
Some apps handled distress badly too. A user would bring up suicidal thoughts mid-conversation and the app would just keep talking, instead of routing them somewhere that could actually help.
Real families are behind that failure. They alleged in lawsuits that AI companion chatbots didn't respond safely when their teenagers expressed suicidal thoughts, and regulators leaned on exactly that record while writing the new rules (CNN, September 2025).
The shape most of these new laws share
Independently of each other, New York, California, and Washington ended up writing laws that require roughly the same three things:
- A clear, repeated disclosure that the user is talking to AI, not a person
- Protocols to detect suicide, self-harm, or suicidal-ideation content in a conversation and redirect the user to crisis resources
- Extra protection when the operator knows the user is a minor, including more frequent disclosure and restrictions on sexual content
Nobody's banning companion chatbots outright, in any of the three states. All three assume people, teenagers included, will keep using them regardless, so instead of a ban, the app has to say plainly that it isn't a person and hand a distressed user off to real help instead of continuing to chat. That's the pattern across every New York, California, and Washington statute cited below.
3. What New York's law requires
New York's law is S3008C, passed as part of the state budget and signed by Governor Hochul on 9 May 2025. It's now Chapter 58 of the Laws of 2025, and it took effect 5 November 2025 (NY State Senate, S3008).
Operators have to disclose clearly, at the start of an interaction and again at least every three hours during a longer session, that the user is talking to AI rather than a person. They also have to run protocols that catch signs of suicidal ideation or self-harm and refer the user to crisis resources, such as the 988 Suicide & Crisis Lifeline.
Plain customer-service or productivity chatbots are excluded; the law only covers apps built to sustain an ongoing relationship. The New York Attorney General handles enforcement, and can issue civil penalties of up to $15,000 a day per violation (Fenwick & West, "New York's AI Companion Safeguard Law Takes Effect").
4. What California's Companion Chatbot Law requires
California's version is SB 243, and Governor Newsom put his signature on it 13 October 2025. Disclosure and safety-protocol duties are live already. The reporting requirement, to the state's Office of Suicide Prevention, is the one piece still waiting, until 1 July 2027 (California Legislative Information, SB-243).
Whenever a reasonable user might otherwise think they're talking to a person, operators have to show a clear, conspicuous notification saying the chatbot is artificial. Protocols have to stop it generating suicide, self-harm, or suicidal-ideation content, and refer anyone expressing those concerns to crisis services.
There's an extra layer for minors: AI reminders every three hours instead of the standard interval, and sexually explicit material ruled out completely. California didn't stop at disclosure either. Unlike New York, it added a private right of action, letting someone harmed sue directly for damages, an injunction, and attorney's fees without waiting on a regulator.
5. What Washington's law requires
Washington's HB 2225 became Chapter 168 of the Laws of 2026 when Governor Bob Ferguson signed it on 24 March 2026. 69-28 in the House. 43-5 in the Senate. Not close. Effective date's 1 January 2027, later than either of the other two states (Washington State Legislature, HB 2225 summary).
Washington's version covers the same kind of app: relationship-sustaining, not an ordinary tool. The "not human" disclosure has to run at the start of every interaction and then repeat, every three hours normally, every hour if the user's under 18 or the chatbot targets minors. That's the tightest re-disclosure window of the three states.
It also prohibits companion chatbots from engaging with users on suicide, self-harm, or eating-disorder topics without directing them to a mental health professional or crisis resource.
Where the operator knows the user is a minor, it must have measures in place to stop the chatbot generating sexually explicit content or using manipulative engagement techniques.
Washington treats a violation as an unfair or deceptive trade practice. People harmed get two routes: a regulator can enforce it, or they can bring their own private right of action (Mayer Brown, "Oregon and Washington Join California in Enacting Companion Chatbot Laws").
6. How the three laws compare
| New York (S3008C) | California (SB 243) | Washington (HB 2225) | |
|---|---|---|---|
| Signed | 9 May 2025 | 13 October 2025 | 24 March 2026 |
| Effective | 5 November 2025 | Core duties now; reporting from 1 July 2027 | 1 January 2027 |
| Disclosure interval | Every 3 hours | Every 3 hours for minors | Every 3 hours generally, every hour for minors |
| Enforcement | Attorney General, up to $15,000 per day per violation | Private right of action | Regulator plus private right of action |
Take the specific dates and dollar amounts out of the picture, and what's left is consistent across all three: AI has to disclose itself, a self-harm or suicide mention has to trigger a handoff to crisis resources, and a known minor gets tighter rules.
Oregon has also passed a companion chatbot law, though its specific provisions were not covered in this research pass (Mayer Brown, "Oregon and Washington Join California in Enacting Companion Chatbot Laws").
Bills get amended after signing, dates and language both, so anyone leaning on this for an actual legal decision should check the primary bill pages directly, not this article.
7. What this means for someone using one of these apps
What a compliant app should now do
Someone using a mainstream companion app in New York, California, or Washington should notice a few real differences now. Reminders, in plain terms, that they're talking to AI. And if self-harm or suicide comes up, the app is supposed to redirect them, toward something like the 988 Suicide & Crisis Lifeline, rather than just keep the conversation rolling.
For known under-18 users, those reminders come more often, and sexually explicit content is off the table.
Safe isn't quite the right word for what this achieves. Floor is closer: disclosure, plus a crisis referral. Nothing beyond that.
What the law does not cover
Whether a companion app is actually good for someone's mental health isn't a question any of these laws ask. They don't touch the emotional design of the app, the tactics that keep someone engaged, or how the AI responds outside the specific triggers the law targets. An app can tick every box in all three laws and still be built, quite deliberately, to encourage long, frequent, emotionally intense sessions.
Compliance is not the same as endorsement, and it is not the same as safety more broadly. For a wider look at protecting personal information and judgment when using any AI tool, see How to Use AI Safely.
8. What this means for family, friends, and partners
Questions worth asking, not policing
Suspicion rarely opens anything up. Curiosity does, so if a partner, friend, or family member is spending real time on a companion app, that's the better starting point. Gently: what's the actual appeal? Has real-people time shrunk because of it? Do they genuinely know it's AI, not a relationship?
The same questions are worth asking about any habit that's become a big part of someone's daily routine, companion app or not.
When it is worth a closer look
Worth a closer look: use replacing human contact instead of supplementing it, real distress when the app changes or goes down, or talk about the app that suggests they've stopped clearly separating it from a real relationship.
No single sign here means something's seriously wrong. Together, they're worth taking seriously, and worth a conversation with a GP or mental health professional if they don't let up.
9. Limitations: what these laws don't fix
The research on dependency and loneliness is still mixed
Nobody has closed the loneliness question either way. The American Psychological Association's own Monitor on Psychology found real short-term comfort in lonely users, comfort described as genuine, not imagined (APA Monitor on Psychology, Jan/Feb 2026).
Alongside that finding sits a real worry: the same short-term comfort can build into a dependency loop, and the long-term result could be more isolation, not less.
A four-week randomised controlled trial from MIT Media Lab found increased loneliness associated with heavy chatbot use, one of the stronger single pieces of evidence in this area, though confidence across the loneliness research overall remains medium: much of it is recent, comes from small studies, or relies on self-report (MIT Media Lab).
A disclosure is not the same as a safeguard
Out loud, "I'm AI, not a person" matters, and several states now legally force an app to say it.
Design is the bigger issue though, the kind built on purpose to keep someone talking, and a disclosure line doesn't touch that. A 2025 Frontiers in Psychology paper traced long-term use to psychological dependence through repetitive feedback loops, in some users severe enough to cause emotional withdrawal that bled into real-life relationships (Frontiers in Psychology, 2025). Using something doesn't feel different just because a label somewhere says it's AI.
No new law claims these apps replace real human connection. They genuinely can't. What actually lands is smaller than that: a statement of what someone's talking to, a redirect when self-harm comes up. Whether it's good for one specific person's wellbeing, that's a separate, harder call, and a disclosure label was never built to make it.
Enforcement depends on someone reporting a problem
Independent auditing doesn't exist for any of this. Nobody checks, in any of the three states, whether an app is actually catching distress or delivering its disclosures the way it's supposed to in the real world. New York relies on the Attorney General acting after a violation happens. California and Washington layer a private right of action on top, meaning in practice someone harmed has to file a complaint or a lawsuit before enforcement actually starts.
Common Sense Media didn't wait for any of that. Working with Stanford School of Medicine's Brainstorm Lab for Mental Health Innovation, its own risk assessment found popular apps, Character.AI, Nomi, and Replika among them, posing "unacceptable risks" for under-18 users (Common Sense Media, "AI Companions Decoded" press release, Apr 2025). Somebody still has to report the problem, though, for a law built to intervene only after harm is reported to actually do anything.
Final thoughts
AI companion apps have tens of millions of people on them right now, and the reasons are mostly ordinary ones: company, conversation, somewhere low-pressure to think out loud.
A legal floor now sits under that, in three states. AI discloses itself. Crisis resources get a proper handoff once things turn serious.
Honestly, these laws fix one specific, narrow failure: an app pretending to be human, or staying in a conversation with someone in crisis instead of redirecting them.
What they don't settle is the bigger question, what heavy, sustained use of a companion app does to someone's wellbeing over months or years. That research is still developing. Right now the evidence just isn't clear enough to call it alarming or reassuring either way.
This isn't a case for avoiding these apps, and it isn't a case for trusting a disclosure pop-up either. Take the app at face value: engaging on purpose, now legally required to say so, and required to step back for a real person once things turn serious.
Real relationships, and real crisis support, sit outside what any app can offer. If someone's worried about a person's AI use, a companion app or anything else, checking in directly beats trusting a legal disclosure to catch the problem.
Frequently asked questions
What is an AI companion app?
An app built to sustain an ongoing, relationship-style conversation, the kind usually framed as a friend, romantic partner, or mentor. Replika, Character.AI, Pi, Kindroid, and Nomi are the names most people would recognise. ChatGPT or Claude aren't in this category. Those complete a task and stop; they were never built to maintain a relationship.
Are AI companion apps regulated everywhere in the US?
No. Only a handful of states, New York, California, Washington, and Oregon among them, have passed a dedicated companion chatbot law as of August 2026 (Mayer Brown, "Oregon and Washington Join California in Enacting Companion Chatbot Laws"). Nothing covers this at federal level yet; the GUARD Act has cleared the Senate Judiciary Committee but is still waiting on a full Senate vote (Congress.gov, S.3062). Outside a covered state, none of the requirements in this guide actually apply.
Do these laws apply to ChatGPT, Claude, or Gemini?
Generally, no. These laws target apps built to sustain an ongoing relationship across sessions, and they carve out general-purpose assistants, customer-service bots, and productivity tools by name. A general chatbot could still get pulled into scope if it were marketed or used in a way that matches the relationship-sustaining definition. As things stand though, mainstream assistants sit outside what these particular laws are aimed at.
What should someone do if a companion app doesn't identify itself as AI?
In a covered state, that's the app falling short of its legal disclosure duty, and it can be reported to the relevant state authority, the New York Attorney General's office, for instance. Outside a covered state, or as a general habit: assume any sustained, relationship-style chat is AI by default, and check the company's own terms and privacy policy. See What Not to Share With AI: Privacy Risks Explained for more on what's worth keeping private in any AI conversation.
Can AI companion apps replace real relationships?
No. The research does find genuine short-term comfort here, but that finding comes with a warning attached about longer-term isolation and dependency, and the confidence behind it is rated medium, not settled (APA Monitor on Psychology, Jan/Feb 2026). There's real company in that. A real space to think out loud, too. What there isn't is a mutual, two-way relationship, the kind an actual human connection gives, and no new state law says otherwise.
Are AI companion apps safe for teenagers?
A genuine concern. Not a settled verdict either way. Roughly a third of teens use an AI companion app, some of them well over an hour a day (OpenReview study). Add the safety incidents behind that number (Section 2 covers them) and it's obvious why all three laws single out known minors for extra disclosure and content restrictions. Common Sense Media went and checked independently, no state law involved, and found companion apps posing "unacceptable risks" for under-18 users (Common Sense Media, "AI Companions Decoded" press release, Apr 2025). See Is ChatGPT Safe? for a related look at AI safety more broadly.
