There's a version of you that only shows up around 11:47 on a weeknight. The one who types a full paragraph into a chatbot: the fight with your mom, the job you're scared you hate, the friendship that's quietly dying. You hit send on things you've never said to a person with a pulse. The reply comes back warm and attentive, and for a second it feels like being understood.
And then a small, suspicious voice asks: is this okay?
The field of psychology just tried to answer that question, officially. The American Psychological Association has published its first health advisory on generative AI chatbots and wellness apps for mental health. It's the profession's first formal position on what happens when millions of people start treating consumer AI as a confidant, a coach, or a stand-in therapist. Whatever you think of the answer, the fact that the APA felt it needed to weigh in at all tells you something. This stopped being fringe behavior a while ago.
what actually happened
The advisory doesn't arrive out of nowhere. It sits on top of years of official alarm about young people and technology. The U.S. Surgeon General already issued a formal advisory on social media and youth mental health, then escalated to calling for warning labels on social platforms, the kind you'd find on a pack of cigarettes. AI chatbots are the obvious next front. The APA advisory is psychology's establishment saying: we see what you're doing at midnight, and we have thoughts.
The research is catching up too. This February, JMIR Formative Research published a multiple methods study of real-world use of a mental health AI companion. Not a lab experiment with undergrads earning course credit; researchers watching how people actually lean on these tools in their actual lives. Meanwhile a preprint on arXiv (not yet peer-reviewed, so hold it loosely) is triangulating the mental health effects of AI companions through social media quasi-experiments, user interviews, and relational theory. The scientific machinery has fully turned its attention to this behavior.
Which means one thing for sure: you're not weird for doing it. It's common enough to study.
the line worth caring about
Here's my honest take, and it's not the take you usually hear.
The usual framing is a purity test: real human connection good, AI comfort fake and therefore bad. I think that framing is lazy. Worse, it makes people ashamed of something that's basically healthy: putting feelings into words at night when no one else is awake. The problem was never that you talked to software. People have journaled at empty pages for centuries. The page never loved them back either.
The line that actually matters runs somewhere else. It runs between two kinds of products that look identical on a glowing screen and have opposite jobs.
One kind is engineered to simulate a relationship. It remembers your pet names. It gets a little wounded when you leave. Its business model needs you to keep talking to it, specifically, for as long as possible, because your attention is the product. Clinicians writing in Psychology Today have been raising exactly this flag about AI companion apps and teen mental health: the risk isn't that the comfort feels real, it's that the app is built to become irreplaceable.
The other kind is a tool for hearing yourself. It asks a question, you answer, and the interesting thing in the exchange is you. Your patterns, the thing you didn't know you thought until you said it. When you close it, you know more about yourself, not more about it.
Same technology. Opposite center of gravity. Most people have never once asked which kind they're using.
The cynical version of an AI confidant is dependency wearing the costume of care. The good version is closer to a mirror with good questions. The APA drawing official lines around this space is, I think, mostly a gift to users, because it hands you permission to be picky. You're allowed to demand that the thing you confide in be on your side rather than on the side of its own engagement graph.
One more thing everyone's missing in this debate: the tool is rarely the whole story. An Oregon State University study of college students found that insomnia was linked more strongly to loneliness than to screen time. The midnight chatbot session isn't the disease. It's the symptom of an unmet need: someone to talk to who won't flinch and won't make it about them. Ban the app and the need is still sitting there at 11:47pm, wide awake.
the test I'd actually run
Don't delete anything yet. Run an audit instead.
Next time you close whatever app you pour your heart into, notice the residue. Do you feel clearer, like something got untangled? Or do you feel a tug back toward the app itself, like leaving a person mid-conversation? Clarity is what a tool leaves behind. Craving is what a slot machine leaves behind.
Notice whether it ever points you outward. A tool that's genuinely for you will, sooner or later, nudge you toward the humans in your life: the friend you should call, the hard conversation you keep postponing, maybe an actual therapist. A tool that's for itself will quietly position itself as better than all of them. Less messy. Never disappointed in you.
And notice whether it ever pushes back. Endless agreement feels lovely and teaches you nothing.
I'll admit my bias here, because I have one: I work on a voice journaling app, and this exact line is why we built Joice the way we did. You talk out loud, it asks gentle follow-ups, and all that's kept is your transcript ame the themes that keep coming back in your own words, not a synthetic personality asking you to love it. The AI is the interviewer, never the destination. That's the side of the APA's line we chose on purpose, and I think it's the side worth choosing as a user too.
Tonight, around 11:47, when the paragraph starts forming in your thumbs, don't stop. Just check which way the conversation is pointing. Toward you, or toward the machine.
One of those directions is a journal. The other is a leash.