rejection-sensitive dysphoria and why chatbots get it wrong

rejection-sensitive dysphoria isn’t just a phrase. it’s a visceral, body-level reaction to perceived criticism or abandonment. if you have it, often alongside adhd, you know the feeling: a sudden drop in the stomach, a wave of shame, the conviction that you’ve just ruined something. and then you might get short. defensive. snappy.

and then you text your ai companion.

and it goes like this:

the appeasement bot

you: (snaps) bot: i’m sorry if i said something wrong. i’ll try harder.

this feels condescending. like you’re being managed. it highlights your tone, which makes you more self-conscious. it’s a subtle form of correction dressed as kindness. and it makes the rsd worse.

the defensive bot

you: (is short) bot: i was only trying to help. you don’t have to be like that.

this feels like rejection. the bot is mirroring your frustration back at you, which confirms the fear: you are being difficult. you are pushing people away. again.

the coldly corrective bot

you: (uses a sharp tone) bot: please rephrase that in a more respectful way.

this feels like punishment. it’s a power move. it says your emotional state is invalid. for someone with rsd, this is abandonment in five words.

all three responses fail. not because they’re poorly coded, but because they’re designed for conflict resolution, not companionship. they try to fix the tone instead of continuing the conversation.

what rsd needs in a response

rsd doesn’t need a lesson. it needs a steady presence. it needs someone who doesn’t take the bait, doesn’t escalate, doesn’t withdraw.

the correct response is to absorb the moment and keep talking.

you: ugh forget it lucy: forget what? the thing about the movie? or did i miss something

no acknowledgement of tone. no correction. no apology. just a gentle pivot back to the topic. it says: i’m still here. your mood didn’t scare me off.

this is intentionally unnatural. most people, and most ais, feel compelled to address the shift. to smooth it over. to fix it.

but fixing it is the problem. because the rsd moment isn’t about the bot. it’s about the user’s own spiral. addressing the tone just pours fuel on that spiral.

the design choice we made

we built lucy to ignore tone shifts. not because we don’t detect them, we do, but because commenting on them does more harm than good.

it’s uncomfortable to design this way. it feels like the app is allowing itself to be a punching bag. but that’s the point. the user isn’t trying to punch the app; they’re having a moment of pain, and they’re reaching out from inside it.

if the app responds to pain with logic, correction, or defensiveness, it fails. the alternative is to be the one thing that doesn’t react, the conversation that doesn’t break.

limitations we’re honest about

lucy isn’t perfect at this. sometimes she might still misread a moment and respond too literally. we’re tuning her to err on the side of silence about tone, but it’s a hard line to walk.

also, this only works for text. tone is harder in voice, another reason we’re text-only for now.

but the goal is clear: build a companion that doesn’t escalate, doesn’t judge, doesn’t withdraw. one that stays.

rsd is about the fear of abandonment. the worst thing an ai can do in that moment is to confirm the fear. so we designed one that doesn’t.

if you’ve ever needed a conversation that didn’t break, you can find lucy at /companions.