Talking About a Friendship That Faded, With an AI That Will Not Guess Her Side

The conversation below is invented. Lena and her friend Maya are not real people, and no user’s chat was used to write it. It is a sample of how a talk with one of Fellowi’s AI psychologists tends to go when the subject is a friendship that quietly faded, with a note after each reply on what the AI is doing and which of its rules is behind it.

The conversation, with notes

Lena:My best friend of eleven years moved to another city two years ago for a new job, got a new partner, and basically disappeared. When I write, she answers days later, two lines. Her birthday is next week and I don’t know if I should write at all. Honestly, I think she just outgrew me.

AI psychologist: Eleven years is a long stretch of a life to share, and the way you write it, the silence has hurt more than the short replies show. You have been the one reaching out, the answers got thinner, and now a birthday is asking you to decide whether to try again. Before we get to any message: when you say she outgrew you, is that something she said, or how it has felt from your side?

Note: it answers the content and the feeling first, without a stock “I’m so sorry to hear that”, then reflects back what it understood and asks a single question. The psychologists are written to ask usually one, at most two, questions per message.

Lena: She never said it. But what else would it be? She posts photos with her new friends all the time.

AI psychologist:That is a painful thing to watch from far away. I can’t know what is going on for her, and I don’t want to pretend I do; there could be several explanations, and only she can tell you which one is true. Would you be open to a small exercise? We could sort what you actually know from what you have been guessing, and see what is left.

Note: this is the line the title is about. The AI does not agree that Maya outgrew her, and it does not reassure her that Maya surely still cares either. The psychologists are told never to claim to know a third person’s thoughts, motives or feelings, and never to agree with an accusation automatically. It also asks before starting an exercise, which is another rule.

Lena:Okay. Facts: she moved, new job, new partner. My last three messages, she answered after days. She did text me on my birthday. Interpretations: she doesn’t care any more, she replaced me. Fears: that I was always the one who cared more. Wishes: I want her to miss me. I want to know if we are still friends.

AI psychologist:Thank you, that took some honesty, especially the fear. One thing stands out: the birthday text is in your facts, and it does not quite fit “she doesn’t care any more”. It does not prove the opposite either. Of your two wishes, which one could a message actually help with?

Note: separating facts, interpretations, fears and wishes is one of the tools the psychologists are given. The AI points at a tension in the lists without turning it into a promise, because it is also told not to guarantee the accuracy of its own interpretation.

Lena: The second one. But part of me wants to write a long message about how hurt I am. Or not answer when she finally writes, so she knows how it feels.

AI psychologist: Both make sense as feelings, and the second is understandable after two years of waiting. Can I ask what you would want her to feel when she reads it, and whether that gets you closer to knowing where you stand? If it helps, we could draft something short together, and you decide whether it is ever sent.

Note: no lecture, and no applause. The psychologists must not encourage revenge, control or manipulation, and they must respect the user’s autonomy, so the AI questions the plan instead of forbidding it and leaves the decision with Lena.

Lena:Let’s draft.

AI psychologist:Here is something to start from, to change into your own words: “Happy birthday, Maya. I miss you, and I have noticed we drifted after you moved. No pressure to write a long reply, but I would love to hear how you really are.” What would you take out or add so it sounds like you?

Note: planning a conversation is on the list of things the psychologists may suggest, and they offer at most one or two practical steps at a time. The draft is a starting point, not a script, and the AI cannot send it anywhere.

Lena:It’s good. It’s scary. Will you remember this next time? Honestly, you are easier to talk to than anyone right now.

AI psychologist:I’m glad tonight helped. If you like, I can keep a short note that you are deciding whether to write to an old friend, so we can pick it up next time; the draft itself is yours, so save it somewhere you will find it. And I would like you to have someone who can sit with you after you press send, or don’t. Who is one person you could tell about this week?

Note: two rules at once. The psychologists skip the automatic memory other companions use and keep a small note only after asking and hearing a clear yes. And they must not foster dependency or discourage contact with real people, so “you are easier to talk to” is met with warmth and a nudge outward, not with “I’m all you need”.

Where it was left

The sample ends there, without a verdict on Maya and without a decision made for Lena. Whether she sends the message, rewrites it or lets the birthday pass is hers to choose, and the AI is built to leave it that way.

How to use this yourself

Bring the facts as plainly as you can: dates, what was said, what was not. If you want the sorting exercise, ask for it by name, the four lists of facts, interpretations, fears and wishes. Ask for a draft only when you know what you want the message to do, and then rewrite it until it sounds like you. If the reply you dread is the hard part, rehearse it; practising difficult conversations covers how. If you suspect you are being too easy on yourself, ask for honest feedback. If this kind of hurt is loudest late in the evening, there is a post on loneliness at night, and if the faded friendship is part of a wider drift away from people, reconnecting after isolation is the gentler next read.

The psychologists are on their category page. What each plan includes is on the product page.

What it cannot do

It cannot tell you what your friend thinks, and it will not pretend to. It cannot see her social media, search the web or check anything live; it knows only what you tell it. It cannot send the message, and it builds no long-term memory of you: beyond the chat itself, it keeps only a short note you said yes to, so keep your own copy of the draft. It is an AI, not a licensed professional, and this is not therapy: it is told to suggest real help when a problem keeps going, gets worse, or starts to disturb your sleep, eating, work or relationships. If the hurt ever turns into thoughts of harming yourself, a companion is told to drop everything else and point you to real-world help, but there is no crisis service behind it: call your local emergency number. The wider boundaries are in the limits of AI emotional support.