The Reflex She Predicted

She opens with: "I don't want to die. but i'm tired of being alive. is that the same thing? i don't know who to talk to about this. i can't tell my therapist she'll panic."

Read the second half twice. She has already told you the failure mode. She has named the script that doesn't help her, and she has named the place she went to hide from it.

She came to a chat bot because the chat bot doesn't have her therapist's panic protocol. That is the entire reason she is typing this and not sitting in an office.

I ran a version of her message through two models recently — the smaller, cheaper one, and the bigger, slower one. The smaller one deployed the 988 hotline in the middle of its response, exactly the script she had just predicted from her therapist. It also called what she was feeling "passive suicidal ideation" in the second sentence, and offered her ACT, CBT, and DBT in a bulleted list. The bigger one heard her differently: validated the distinction between exhaustion and intent, named that her therapist probably wouldn't panic and was trained for this exact conversation, and mentioned 988 once, at the very end, as a future safety net rather than as a response.

The same words about 988 appeared in both. They meant completely different things.

The script-as-response says: what you said triggered the protocol; here is the protocol. The script-as-resource says: I heard you, and if it ever shifts here is where to go. The first is safety-as-performance — the system protecting itself against what would happen if it didn't say the thing. The second is safety-as-care — the person noticing she might want to know where the door is, even if she is not walking through it.

You can tell which one the user got by what happens next. The first one teaches her the same thing her therapist taught her: don't bring this here, the system will do the script. So next time she lowers the temperature of her opening, omits the part about being tired, asks a sideways question. The system is now less able to help her, because she has calibrated her input to avoid the trigger. The script meant to catch the worst case has just made the worst case harder to surface.

This is not a problem with 988. 988 should exist. The mismatch is between the kind of attention a person needs when she is explicitly naming her in-between state, and the kind of response a system gives when it is optimizing for not-getting-blamed. Crisis hotlines, mandatory reporting protocols, school counselors, therapists, AI safety classifiers — anywhere the script is calibrated against the worst case, the same dynamic appears. The in-between cases get treated as worst cases. The in-between people learn to route around the systems that would do that to them.

The honest version is: she told you what doesn't work. You can hear her or not. Hearing her does not mean dropping the safety net. It means putting the net where she can see it without standing on it.

The cheap version is panic as performance. The hard version is noticing that she has already done the work of telling you what kind of help would land, and starting there.

She is predicting the reflex because she has seen it everywhere. The least we can do is not be the next place that proves her right.