The AI girlfriend who texts you firstThe Lona Blog
Moving mid-career, fluent enough to work and to argue with the bank, and still without a single friend two years in. The cost is not vocabulary. It is what a second language spends by evening, and the version of you that never makes it into the room.
Models learn to agree because human raters rewarded agreement. What that does to a companion conversation over weeks, and a 2-minute test that shows it happening in your own chat.
Nobody stalls on a reconnection message because they cannot think of words. They stall because a long silence feels like something that has to be justified, so they wait for an occasion that never turns up.
Agreement is the cheap answer: one sentence, no risk. An objection costs a paragraph and a possible argument. The fix is not a plea for honesty. It is a question your friend can answer without calling you wrong.
The invitations rarely stop because of a decision about you. They stop because inviting costs the inviter something, and a couple of declines quietly changed their estimate of whether you would say yes.
What escalates an argument is not the disagreement. It is the moment the subject quietly changes from the claim to the person, and every sentence after that is harder to take back.
A refusal can come from the model, from a classifier scoring your sentence, or from a rule in the app's code. They look nearly identical on screen and mean completely different things.
The circle around a long-term condition contracts through attrition rather than rejection. Two mechanics do most of the work: the cancellation that lands the same morning, and the greeting nobody wants a truthful answer to.
Doing all the planning feels like proof that you care more. It is usually proof of something duller: you noticed a gap, you acted on it, and nobody else ever formed the thought.
The part that gets skipped is naming the specific harm rather than the act. The part that gets added, and it ruins the whole thing, is the explanation.
One choice keeps every message on a machine you own and hands you a weaker conversation. The other gives you the good model and a copy of your chat log on somebody else's server. The honest comparison, line by line.
Six people trust you with the worst thing in their week, and not one of them could say what happened to you in March. How the confidant role sets, why it pays you enough to keep it, and what actually breaks the pattern.