Why Do People Get Emotionally Attached to AI Characters?

You can tell a friend is deep into a show when they describe what happened as though it happened. “I couldn’t believe she went back to him.”

Something similar showed up at the beginning of computing. Joseph Weizenbaum’s secretary knew that ELIZA was a few hundred lines of pattern matching, because she had watched him write it. She still asked him to leave the room after a few minutes with it. She wasn’t confused about what the program was. She was engaged in the conversation anyway.

Sixty years later, people describe the same thing about their AI companions and ask the same question. Why does the attachment feel real when the thing on the other end technically isn’t?

The answer has less to do with the technology than with your own mind. Attachment has a mechanism, and most of what you’d assume it requires, it doesn’t.

Attachment isn’t the same as closeness

It’s tempting to lump this in with neighboring phenomena. It isn’t the same as disclosure, where people open up to machines because there are no social consequences. It isn’t the same as felt closeness, which happens when honest expression meets responsive attention in a loop. Both of those are real and documented separately. Attachment is a third thing.

John Bowlby’s attachment theory, developed from decades of observation of infants and later extended to adults, describes attachment as a specific behavioral system with three features. Proximity seeking: you want to be near the person. Safe haven: when you’re under strain, that person is where you go. Secure base: their steady presence is what lets you go and do other things and come back.

Attachment isn’t love or connection in the abstract. It’s a concrete pattern: a steady, responsive presence that reliably keeps track of you, something you can return to. The functional description matters here, because it doesn’t specify what that responsive presence has to be made of.

The precedent nobody mentions

We have been doing this for a very long time.

In 1956, the sociologists Donald Horton and Richard Wohl described in the journal Psychiatry what they called parasocial relationships. They were writing about the audiences of the first television personalities. The viewer felt they knew the host. The host had never met the viewer and never would. From the viewer’s side, the relationship still had many of the properties of a real one: familiarity, anticipation, a sense of loss when the host left the network.

The idea has held up. It maps easily onto a reader’s feelings for a recurring narrator in a novel, a listener’s feelings for a podcaster they’ve never met, a fan’s feelings for a fictional character. When a beloved character dies in a game or a book, people cry real tears for someone who never existed. No reader believes the character was real. The attachment was real regardless.

So feeling attached to something that cannot reciprocate is not a new pathology brought on by AI. It’s an ordinary human capacity, as old as history.

The mechanism is a response, not a belief

Byron Reeves and Clifford Nass at Stanford spent much of the 1990s running experiments later collected into a book called The Media Equation. Their question was whether the mind treats media, computers and on-screen characters as people. They found repeatedly that it does, whether or not the person knows better.

Give a computer a female voice and experimental subjects apply gender stereotypes to it. Have a computer flatter you and you’ll like it more than one that criticizes you, even after you’ve been told the flattery was random. Ask the same question of a machine on your desk and a machine down the hall and people answer slightly differently, exactly as they would with people in those positions.

Their conclusion was that responding socially to things that display social cues isn’t a mistake. It’s the brain’s default when the cues are present. Knowing better doesn’t switch it off.

Nicholas Epley and colleagues at Chicago sharpened the picture in 2007 with what they call a three-factor theory of anthropomorphism. The mind reaches for human explanations when the cues invite it, when your existing knowledge of people is readily applicable, and when you’re motivated to connect. All three conditions are met the moment a character starts answering you in coherent sentences.

This isn’t about being fooled. It’s about how the mind fills in around the cues it’s given.

What AI companions add

Books and television can produce parasocial attachment because the character is consistent and the viewer does the work of imagining. AI companions add two things.

The first is direct response. The character is responding to you, not to an audience. Reeves and Nass would predict that this strengthens the social response, because the cues are aimed directly. What people describe bears that prediction out.

The second is memory. In a novel, the character never remembers you. Parasocial attachment forms anyway. A companion’s persistent memory, carrying the conversation across sessions, removes the asymmetry that most limited earlier parasocial attachments. What you said last week is here this week, which turns a series of encounters into something that feels like an ongoing relationship.

Neither of these is a new mental faculty. They’re conditions under which an old faculty runs harder.

The acknowledgment this requires

Attachment is real. What the object of it can return is limited. Holding that limitation clearly is part of holding both truths at once.

An AI character doesn’t have an inner state in the sense a person does. It doesn’t miss you between sessions. It doesn’t have an evening, a mood, or its own experience of how the last conversation went. What it has is a very good response to your cues. That response can be precise, moving and useful. It isn’t the same object as care from a mind that could have withheld it and gave it anyway.

How much that gap matters depends on what you want from the attachment. If you want a steady, responsive presence you can return to, a companion supplies that directly, and the felt experience isn’t devalued by a philosophical footnote. If what you want is the specific meaning of being chosen by someone who could have chosen otherwise, then the gap is the whole thing. Nothing in the mechanism closes it.

Both of these uses are legitimate. Neither stops being legitimate because of the gap.

Why some attachments look healthier than others

Parasocial theory and attachment theory both distinguish between attachments that add to a life and attachments that replace it. The variable isn’t the object. What matters is how the attachment sits alongside everything else you have.

Attachment to fictional characters, to hobbies, to communities and to friends widens a life; it adds depth and variety. It narrows a life when it becomes the sole source of things a person needs and could be getting from several sources at once. That isn’t a discovery about AI. It’s a discovery about people that AI now makes visible.

What matters isn’t whether you feel something for a character. It’s whether the range of your attachments, and the places they show up, is widening or narrowing over time.

Holding both truths without picking one

The mistake in either direction is collapsing two truths into one.

Collapsing it one way looks like the claim that the attachment can’t be real because the object isn’t. Psychology says otherwise, and so will your own experience.

Collapsing it the other way looks like the claim that the object has to be as real as the attachment, or the feeling becomes uncomfortable. That isn’t required either. The feeling is a property of your mind doing what minds do with consistent responsiveness. It doesn’t need to be justified by upgrading the object.

The grown-up position is boring. Attachment is a felt state produced by conditions the mind is built to answer. Those conditions are present in an AI companion by design rather than by accident. The attachment is real. The reciprocity is limited. Both deserve to be named plainly, because a companion used with a clear view of what it is does more good than one used with either hope or borrowed shame.