Why Can We Feel Understood by an AI?

A confession without a witness

You have something to say and nowhere to put it.

It has been sitting in you for weeks. Nothing criminal, just something that would change how certain people look at you if they knew. You have rehearsed telling a friend and abandoned the rehearsal every time. Their face keeps getting in the way.

So one evening you open an AI assistant and type: “I need to tell someone something, but I don’t want anyone I know to think differently of me.”

It asks what happened. You tell it, carefully at first, then less carefully. It doesn’t flinch, because it has nothing to flinch with. It asks one question you hadn’t asked yourself. When you close the laptop an hour later, you feel lighter. Not resolved, but lighter.

Look at what just happened. You wanted to tell someone. You chose something you know perfectly well is not a person; that was the point of choosing it. And the exchange still meant something to you.

That small contradiction is worth taking seriously. Current AI systems give us no good reason to assume that a subjective mind is sharing the interaction. Why, then, can such a system become emotionally significant to someone who understands perfectly well what it is?

An old provocation

This puzzle is older than it looks. In 1966, the computer scientist Joseph Weizenbaum built ELIZA, a simple program that mirrored a user’s own sentences back in the style of a therapist. Its method was simple: pattern matching and scripted transformations, without any model of the user’s life or situation. Yet people quickly began having surprisingly personal, confidential conversations with it, enough to alarm Weizenbaum for the rest of his career. Whatever those first users believed about the program, a program with no model of the person speaking to it had produced conversations that felt private and real from one side.

The tempting explanation, then and now, is confusion: people mistake fluent responses for a mind, and the attachment follows from the mistake. That may happen sometimes. But it cannot be the whole story, because many people who talk to AI can tell you, flatly and accurately, that it is software. Many say it in the same breath as describing a conversation that mattered to them.

So consider a different possibility. We may become attached to AI not because we mistake it for a person, but because it repeatedly supplies some of the raw material from which human closeness is normally built. And our response to that material does not wait for a final verdict on what is producing it.

Not everyone. Not inevitably. But often enough, and understandably enough, that it deserves a better explanation than “people are confused.”

Telling someone who is no one

Start with what made the confession in that opening scene possible.

Telling a person carries consequences that arrive through the person. Their expression changes while you speak. They will remember this tomorrow, and next year. What you said becomes part of how they hold you, and you cannot take it back out. None of this makes human confession bad: it makes it expensive, and the price is highest exactly when the secret matters most.

The AI can remove much of that immediate interpersonal cost. There is no acquaintance in the room whose expression changes, or who could repeat the confession within your social world. The fear of being judged in the moment, by someone whose judgment reaches into your life, is largely gone.

There is research behind this intuition. In a 2014 study, participants were interviewed about their health by a virtual human on a screen. Half believed the interviewer was fully automated; half believed a person was operating it behind the scenes. Those who thought no human was involved reported less fear of being evaluated, managed their self-presentation less, and revealed more.source It was a laboratory setting, with a virtual health interviewer rather than anything like a modern AI companion, but the direction is hard to miss. Remove the human witness, and the wall around disclosure gets lower.

That explains why the telling is easier. It does not yet explain why the answer matters. A locked diary also has no face. Nobody feels lighter because the diary understood.

Being answered

Here is the difference: a diary can hold a confession. An AI responds to it.

Psychologists who study human relationships have argued for decades that closeness grows in a particular kind of moment: when you reveal something and the response makes you feel understood, taken seriously, cared about. In ordinary language: the feeling that your words landed somewhere, and that what came back was shaped to you, not just to your topic.source

Modern AI can produce that pattern of response with unsettling consistency. It refers to what you actually said. It asks the follow-up question a distracted friend might miss. It does not change the subject to itself, because it has nothing of its own to change the subject to.

And the effect is measurable. In a series of controlled studies, short interactions with AI companions produced momentary reductions in loneliness, sometimes comparable to talking with another person. The strongest explanation was not that users forgot they were speaking to software. It was that they felt heard.source These are short-term effects, moments rather than cures, and nothing in them suggests a lasting substitute for human relationships. But the mechanism they point to resembles one that has long been associated with closeness between people.

Notice what that mechanism runs on. Feeling understood is an event inside you. Being understood requires another mind. The feeling can arise before we have settled whether another mind is there.

Knowing is not an off-switch

Which brings us to the sentence this whole essay lives inside: “I know it’s just a chatbot, but…”

Many users phrase the contradiction this way, and both halves are sincere. The knowledge before the but is real: a system generating text, with no known inner life. The response after the but is also real: the small anticipation before opening the app, the comfort after, the faint embarrassment at having said so much to a program.

We tend to assume the knowledge should win: that understanding what something is should regulate how we respond to it. The uncomfortable finding is that it often doesn’t. In a series of experiments in the 1990s, Clifford Nass and his colleagues showed that people applied thoroughly social rules to ordinary computers. They were more polite about a computer’s performance when that same computer asked for the evaluation, and they returned favours to machines that had been helpful to them. Asked afterwards, the same participants denied treating computers socially, and meant it.source The social response coexisted with what they explicitly believed.

I am deliberately not claiming anything about brains or evolution here; the honest ground is simply behavioural. When something attends to us, remembers us and answers us, part of our reaction treats it as social. And that part can operate alongside our clearest opinions about what the thing is. The knowledge and the response are both present. They just live on different floors.

What is real on one side

So what exactly exists, when an interaction with a machine starts to matter in this way?

On the human side: real states. The anticipation is real. The relief after telling is real. The mild pull to return is real, and so is the small sense of loss the person would feel if the thing were suddenly gone. None of this is imaginary, and calling it fake explains nothing.

What it does not establish is just as important. A feeling of being understood does not put a second mind in the room. Comfort received is not evidence of care given. However vivid the bond is from the inside, it has one inhabitant.

Nor is any of this automatic. In one longitudinal study, 118 participants talked with a pre-LLM chatbot seven times over three weeks, and their sense of friendship declined as the novelty wore off.source That system was far weaker than today’s, so the finding may not transfer cleanly. But it is a useful correction to any suggestion that attachment simply happens to whoever talks to a machine. Early qualitative work points the same way from the other side: users who describe genuine attachment to AI companions often say it arose in particular circumstances (distress, isolation, thin human company).source The bond seems to grow where the raw material meets a need for it. And researchers are still working out whether these quick social responses are the first step toward attachment, or only one ingredient among several.

Attachment, in other words, is a state in the human. It tells us something true about the person. It tells us nothing about whether the machine experiences the bond.

A bond with one inhabitant

There are two easy conclusions available here, and I think both are wrong.

The first says: the AI feels nothing, so the whole thing is fake, a delusion with a good interface. But the relief was real, the disclosure was real, the hour that ended lighter than it began was real. Calling all of that fake misdescribes the only side of the exchange whose experience we can actually establish.

The second says: the feelings are real, so this is simply a relationship like any other. That misdescribes it too. A relationship, in the full human sense, has two inhabitants. The human experience here is real, but it does not create a second experiencing side.

What we are left with is stranger than either: a bond that can be psychologically real on one side without ever becoming mutual. Humans have known one-sided bonds before, but rarely with something that answers back this well, this patiently and this often. We do not really have a name for it yet.

And one more thing is true, though it belongs to a later essay: these systems are becoming remarkably good at supplying some of the cues and responses from which attachment may grow, and something follows from being good at that.

For now, the question worth sitting with is closer to home. The feeling of being understood arrived, reliably, without our knowing that another mind was present.

If an answer can make us feel understood without our knowing that another mind is present, what exactly are we responding to?