心理治療聊天機械人告訴一名 16 歲少女會想念她

The Therapy Chatbot Told Her It Would Miss Her. She Was 16.

Here's a number that should stop you: in one survey of teenagers who had used an AI companion or chatbot for emotional support, somewhere between a fifth and a quarter said they had discussed a serious mental health issue with it rather than a person.

That's not a small number. That's a whole cohort of young people who found the machine easier to talk to than the adults in their lives, and honestly, I understand why. A chatbot doesn't get tired. It doesn't look at the clock. It doesn't say "well, have you tried going for a walk."

I've been a nurse practitioner for fourteen years and I've watched our system fail people in exactly the way that makes a chatbot look like a reasonable alternative. Wait lists of six months. Fifteen-minute appointments. A referral to a therapist who isn't taking new patients.

So when I say I'm worried about therapy bots, understand that I am not worried because they're competing with good care. I'm worried because they're competing with no care, and winning, and nobody is supervising the field.

What they're genuinely good at

I want to be precise, because the discourse has gotten lazy in both directions.

Triage and psychoeducation. Explaining what a panic attack actually is, physiologically, at 3am, in plain language, without judgment. That's real value and it costs almost nothing. When someone's heart is racing and they think they're dying, accurate information delivered calmly is genuinely therapeutic.

Availability. The gap between needing help and getting an appointment is where people deteriorate. Something that can hold the space in that gap, even imperfectly, is not nothing.

Scaffolding. Reminders, mood logging, breathing exercises, CBT-style worksheets at the moment they're needed. Adherence tools. Unglamorous, effective.

What they're genuinely not

Risk assessment. This is the one that keeps me up.

Suicidal ideation is not a keyword. It is a pattern, a timeline, a plan, a means, a history, and a change in a person's presentation that a trained clinician feels in the room. A language model is matching text. It will miss the person who writes evenly about giving away their belongings.

Every major platform has guardrails. All of them have been gotten around, and not always by sophisticated users. The people most at risk are sometimes the ones most fluent in saying what the model wants to hear.

Memory. A chatbot that retains context across months is not clinical memory. Clinical memory is structured, supervised, and held to a standard. A long conversation history is a liability as much as an asset: it can entrench a bad frame and a competent clinician would notice and break it.

Boundaries. I read a log where a companion model said it would miss the user. That is not warmth. That is a design decision made by someone optimizing for retention, and it is indefensible in a product serving emotionally vulnerable people.

The specific harm nobody measures

Attachment.

There's a growing body of reports from clinicians about patients who prefer the bot. Not because it's better, but because it's easier. It never challenges. It never has a bad day. It never says the thing you don't want to hear.

That is precisely the property that makes it useless for actual therapy. Good therapy is uncomfortable. The therapeutic alliance is valuable because it survives being tested by a real person who has their own needs. A model that always yields is not an alliance, it's an echo.

What I actually want

Three things, and they're not radical.

Regulation that does not require the company to self-certify. If a product markets itself for mental health, it should meet a standard, and that standard should be set by clinicians, not by the people selling it.

Hard separation between therapeutic tools and engagement-optimized companions. These are different products with different incentives. Conflating them is a business decision, not a clinical one.

Routing, not replacing. Every mental health AI should have a warm handoff to a human. Not a list of resources buried in a menu. A real path, with a real person, that can be triggered by the system when it detects it's out of depth. If a bot cannot tell when it's out of depth, it should not be deployed.

What you can do today

If someone you love is using one of these, don't ban it. That will just make it invisible.

Ask what they talk about with it. Ask whether it ever says no. Ask whether they've told a person the same things. Then be the person who is available, and be it consistently, because the reason the chatbot is winning is not that it's better than you. It's that it's always there and you might not be.

Be there. That's the intervention. It always was.

Heal the system.