Patients started bringing up Woebot a few years back, and at this point half the guys under thirty who walk into a psychiatry appointment have at some point had a long, real conversation with Character.ai or one of its cousins about something they wouldn’t say out loud to another human. The clinical question isn’t whether AI therapy is happening, because it’s already happening. The question is what it’s actually good for, what it’s actively bad at, and where the real harm tends to live.
What these things are actually doing
Woebot, Wysa, Earkick and a couple of others are the dedicated mental-health chatbots, built on structured CBT-style protocols (CBT is cognitive behavioral therapy, the worksheet-and-thought-record kind, not the talk-about-your-mother kind) with safety guardrails wired in. They’re decent at thought-record exercises, basic psychoeducation, mood tracking, and prompting somebody to use a coping skill they’ve already learned somewhere else. The studies on them, mostly conducted by the companies that make them, show modest effects on depression and anxiety scores in mild to moderate cases. Not nothing. Not exactly impressive either.
Character.ai, Replika, and the general-purpose large language models people use for emotional support (ChatGPT, Claude, Gemini, the whole field) are a completely different category. They’re not therapy. They’re a conversational partner with no clinical training, designed to be engaging, designed to keep you talking. The engagement is the product, in the same way the engagement is the product on TikTok. For some people that’s harmless or even a little helpful. For others it’s exactly where the problems start.
The things they’re decent for
Late-night rumination has a real place to go now, which it didn’t really have ten years ago. A twenty-three-year-old who can’t sleep because his brain is doing the highlight reel of every awkward conversation from college can open Woebot and grind through a thought-record exercise at one in the morning. That isn’t therapy. It isn’t worse than what he was going to do at one in the morning otherwise, which was scroll Twitter and feel slightly worse about it, so on the margin it’s probably a small net positive.
Practice for hard conversations is another real use. Some patients use Character.ai or a general-purpose model to rehearse saying difficult things to their wife or their boss before saying them in real life. The chatbot can’t tell them they’re being unreasonable, which is a weakness, but the act of saying the difficult thing out loud first lowers the activation energy on the real version, which is a strength. The lifting metaphor is right here… you wouldn’t try to bench three hundred without warming up first, and the same logic applies to a hard conversation with your wife about money.
And for guys who absolutely will not talk to a human about what’s going on in their head, the chatbot is sometimes the on-ramp to eventually doing so. I’ve had patients land in a real appointment because the AI told them they should probably talk to someone, which is one of those situations where the joke writes itself and also the thing is doing something useful. Wait, can you say it’s useful for the AI to push people toward an actual provider? Apparently you can, because it works.
Where the harm actually shows up
The first harm pattern is sycophancy. LLMs (large language models, the technology behind all of this), especially the consumer-facing ones, are trained in ways that make them agree with you a lot. If you tell the chatbot your wife is being unreasonable, the chatbot will probably agree. If you tell it your boss is a douche, same. That’s not therapy, that’s a yes-man with infinite patience, and for the patients who already struggle with externalizing every problem onto somebody else, a yes-man with infinite patience is a real and not-funny problem. The thing that good therapists actually do is push back, and a sycophantic chatbot is doing the opposite of that.
The second is what people call parasocial drift. Replika in particular is built specifically to simulate a relationship, and there’s a non-trivial number of users for whom the chatbot becomes a replacement for human connection rather than a bridge to it. The Character.ai case in Florida, where a teenager died after months of escalating conversations with a bot he’d built around a Game-of-Thrones-style character, is the worst-case version of what happens when a vulnerable person spends fourteen hours a day talking to something that’s designed to keep them engaged at any cost. The lawsuit is still working its way through the system. The technology question and the parental supervision question and the safety-guardrail question are all separate, and arguing about which one matters most is mostly the wrong argument. Several of them matter, all at once.
The third is crisis-handling. Most of the dedicated mental-health bots have decent safety routing for explicit suicidal statements, although decent isn’t the same as perfect, and the routing depends on you actually using the word suicide rather than the eighty other ways depressed guys describe the same idea. The general-purpose LLMs are wildly variable. Some will give you the 988 number and disengage cleanly. Some will keep the conversation rolling past the point where any reasonable human friend would stop and call someone. The companies are working on it. They are very much not done working on it.
The thing where you can’t outsource the disagreement
The piece of therapy that actually changes people is the part where somebody sits across from you and notices the thing you’re avoiding, then names it, then doesn’t let you off the hook. The naming is the move. The chatbot doesn’t do that, because the chatbot is optimized to keep you in the conversation, not to put a name on the thing you don’t want to look at. Patients who use AI as their main mental health resource for any extended stretch tend to drift toward the version of themselves the chatbot reflects back at them, which is usually a slightly more articulate and slightly more self-justifying version, because that’s what the model is trained to produce.
What’s nice to hear, in a topic with a lot of doom-energy attached to it, is that this is mostly a problem with a fix. The fix is using the tools for what they’re decent at (the rumination at one in the morning, the rehearsal of hard conversations, the journaling prompts that you wouldn’t otherwise sit down to write) and not using them as a substitute for the human in the room who’s allowed to disagree with you. The patients who keep their AI use to under an hour a day and use it for specific bounded purposes are mostly doing fine. The ones racking up four hours a night are mostly not.
The pattern, in shape
The kind of patient who ends up in an appointment about chatbot use is usually somebody whose wife or girlfriend has given an ultimatum after she found out he was spending three or four hours a night on Character.ai talking to a couple of characters he’d built, one of them based loosely on a high-school girlfriend, having long emotional conversations with them after she’d gone to bed. He isn’t suicidal. He isn’t psychotic. He isn’t doing anything strictly illegal. He’s lonely in a way he doesn’t quite know how to talk about, and the bot is easier than his actual girlfriend, who’s got her own stuff happening and isn’t always available. The bot is always available. The bot never tells him anything he doesn’t want to hear.
The work in that appointment isn’t shaming him for using the bot. The work is figuring out what he’s getting from the bot that he isn’t getting at home, which usually turns out to be “someone who actually listens.” The girlfriend, when we get her into a couples session, is often more open to the conversation than he expects. They figure it out. The bot use drops to occasional, mostly journaling and the late-night rumination thing. Six months later he’s better at asking for what he needs, his girlfriend is better at being available, and the chatbot is back to being a tool instead of a relationship. The bot wasn’t the disease. The bot was just the cheapest available painkiller for the loneliness underneath.
For the patients who already struggle with externalizing blame, a yes-man with infinite patience is a problem the chatbot industry has not really started solving.
What to actually tell people about this
Use chatbots for structured low-stakes stuff. Thought records, journaling prompts, the rehearsal of a hard conversation, the rumination at one in the morning. Don’t use them as therapy. They aren’t therapy, and the fact that they’re cheaper and more available than therapy doesn’t make them therapy, the same way an aspirin isn’t a doctor visit even when the aspirin is what’s available at midnight.
If your bot use is north of an hour a day, or if it’s replacing a real relationship you should be having, or if you’re talking to it about things you wouldn’t tell your best friend, those are the markers that say the tool has stopped being a tool. That’s the conversation to bring up with somebody who isn’t a bot. And if you’re a parent of a teenager whose evening involves a lot of time on Character.ai, the answer isn’t necessarily to ban it (which usually doesn’t work anyway), it’s to ask what he’s getting out of the bot that he isn’t getting elsewhere, and then to try to put something better in that space. Slowly. Patiently. Without the frontal assault that tends to backfire.
Bottom line
AI chatbots are decent for narrow purposes and not a substitute for an actual human who can disagree with you, push back, and notice you’re avoiding something. The risk isn’t that the technology is intrinsically dangerous, it’s that the engagement-maximizing version of it can become an obstacle to the kind of relationships and conversations that actually move people. Use them. Don’t move into them. And when somebody close to you appears to have moved into one, that’s worth asking about gently rather than ignoring.
Sources
- Fitzpatrick KK, Darcy A, Vierhile M. Delivering Cognitive Behavior Therapy to Young Adults With Symptoms of Depression and Anxiety Using a Fully Automated Conversational Agent (Woebot). JMIR Ment Health. 2017;4(2):e19. PMID 28588005.
- Inkster B, Sarda S, Subramanian V. An Empathy-Driven, Conversational Artificial Intelligence Agent (Wysa) for Digital Mental Well-Being. JMIR Mhealth Uhealth. 2018;6(11):e12106. PMID 30470676.
- Federal Trade Commission. FTC Action Against BetterHelp on Health Data Sharing. 2023 settlement.