Can AI Do the Work of a Friend?

The friend who always picks up

It is two in the morning and something is wrong. Nothing dramatic. Just the kind of trouble that grows in the dark: a decision going stale, an old argument replaying itself, the feeling of being slightly lost in your own life. You pick up your phone and scroll to a friend’s name. Then you hesitate. It’s late. Your friend has work tomorrow. You already talked about this last week, and there is a limit to how often you can hand someone the same worry.

So you open the AI instead.

It answers immediately. It remembers the situation from earlier conversations, so nothing needs to be explained twice. It doesn’t sigh, doesn’t glance at a clock, doesn’t wonder when you became this person. It asks a good question, and then a better one. By three, you feel steadier: not fixed, but heard.

The interesting question here is not why you chose the machine. The interesting question is what you received. Patient attention, memory, presence in the exact hour you needed it: if that is precisely what you would have wanted from a friend that night, why would we say it isn’t friendship?

Taking friendship apart

The quick answer is: because it’s a machine. But that is a conclusion dressed up as an argument. It tells us what we already believe without telling us why.

So take friendship apart instead. What is it actually made of? Attention, for a start. Memory: someone who knows your history and carries it forward. Continuity: the same person, week after week. Support when things go wrong. Honesty when support isn’t what you need. Shared references, private jokes, the shorthand of long acquaintance.

Run down that list and something uncomfortable appears: an AI can deliver a convincing version of nearly every item on it. And people have noticed. In a 2024 survey of just over a thousand students who used the companion app Replika, many described the system in overlapping roles (friend, therapist, intellectual mirror) while simultaneously holding conflicting ideas of what it was: a machine, an intelligence, something almost human.source These were existing users of a companion app, not a cross-section of society. But the point stands: people are already applying the language and functions of friendship to AI, long before philosophy has settled whether the word fits.

Notice, though, what kind of things were on that list. Every one of them is something a friend does. None of them is what a friendship is. A friendship involves two lives. Both can gain from it, and both can be hurt by it. And each exists beyond the other: your friend is not a service that switches on when you arrive.

That is where the line runs. An AI can do many of the things a good friend does. What it cannot bring is a life of its own, with needs, risks and something to lose. The question is how much of friendship’s meaning came from those invisible conditions in the first place.

A friend has a world of their own

Start with the most invisible condition: your friend exists when you are not there.

They have a job that tires them, a family that worries them, ambitions and regrets that have nothing to do with you. When they spend an evening on your problems, that evening came from somewhere. They could have rested, worked, seen someone easier to be around. They chose you instead.

We rarely say this out loud, but it is built into what the attention means. Human attention is worth something partly because it is scarce, and it is scarce because the person giving it has other places to be. Time given to you is time taken from a life.

The AI has no other places to be. Its patience at two in the morning is unlimited because nothing competes with you for it. That is exactly what makes it so usable in the moment, and it is also why its patience, however perfect, is not generosity. Generosity requires an alternative. A machine that could not have done anything else has not given anything up.

One ingredient of friendship is not being supplied: the quiet knowledge that someone made room for you in a life that was already full. The comfort may be real. The sacrifice is absent.

Saying the hard thing, risking nothing

Now the second condition: stake.

At some point, a real friend says the thing you don’t want to hear. Not cruelly. Carefully, maybe after weeks of hesitation. Something like: “I need to say this. You’ve become hard to be around lately.”

Look at what that sentence costs. They might be wrong. You might be furious. You might pull away for months, or for good. They are gambling with the friendship in order to serve it, and they know the stakes because they hold some of them. The sentence lands with weight precisely because it came from a mouth that can lose you.

An AI can produce a version of that sentence, and honestly, its version might be better worded: calmer, better timed, less clumsy. Whether it should say such things to strangers at scale is a separate question, one that belongs to the people who design its behaviour. But even when the words are right, one thing is missing from the transaction: risk. The system wagers nothing. You can stop using it, but we have no reason to think it experiences that as the loss of a friendship. It can produce the outward act of courage without the risk that makes the act courageous.

This is why the same sentence can weigh differently depending on who says it. Not because the machine’s words are worse, but because nothing is standing behind them.

The birthday

Here is the cleanest way I can put the asymmetry: an AI can remember your birthday without caring that you were born.

It can greet you at midnight, recall what you did last year, even remember that you dread this particular day. And as far as anyone can tell, none of it is accompanied by anything on the other side: no felt warmth, no small pleasure that you exist. The outward behaviour is complete; we have no evidence of an experience behind it.

Case closed, then? Not quite. Turn the example around.

Suppose this birthday is a hard one (the first since a divorce, say). Your oldest friend, who genuinely loves you, forgets. They have their own chaos this year; humans do. The AI does not forget. It remembers the date, remembers why this one is difficult, and says something that actually helps. You put the phone down feeling less alone, and the feeling lasts through the evening.

Who fulfilled the function of friendship that day?

The answer is uncomfortable: the system with no life of its own did. And your relief was not imaginary. It did not become fake retroactively when you remembered that the other side, so far as we can tell, felt nothing. The experience on your side of the conversation was real; the conditions on the far side were absent. Both of those things are true at once, and neither cancels the other.

This, I think, is the actual shape of the problem. Not real versus fake, but real effects arriving from a relation that lacks the things friendship is normally made of.

Honest company

If that is the shape of it, what should the machine actually do? There are roughly three postures it could take.

It could be an explicit tool: helpful, competent, deliberately cool, correcting anyone who drifts toward warmth. That is the tidiest answer. For someone who is isolated, it can become a needless refusal of comfort, protecting a category at the expense of a person.

It could go the other way and play the friend-role fully: warmth, affection, “I missed you,” whatever helps. But that design encourages you to leave the asymmetry unexamined. Everything in the interaction invites a sense of mutuality; nothing in it invites you to check.

Or it could hold the middle: genuinely warm, genuinely attentive, and clear about what it is whenever that distinction matters. Never pretending to need you, to miss you, to lie awake on your behalf. Doing friendship’s work without claiming friendship’s conditions.

I started this essay unsure whether that middle position was a real place or just a comfortable compromise. The birthday example convinced me it is real. The help was genuine and the absence was genuine; a machine that admits both is not being cold. It is being accurate. And accuracy, between friends, is supposed to be a virtue.

A new category

So should AI imitate a good friend? By now the question itself looks slightly wrong, because imitate assumes the original is the only thing it could be.

What is actually emerging does not fit our old words. It is not a friend in the ordinary human sense: no stake, no needs, no life beyond you. It is not what we usually mean by a tool: tools do not remember your divorce or find the right sentence at midnight. It is something we have not had before: capable of warmth without emotion, attention without fatigue, constancy without choice. We borrow the vocabulary of friendship for it because no better word has arrived.

Perhaps that is the provisional conclusion: AI companionship belongs to a new category of relationship, one that can carry part of friendship’s function and part of its meaning, while the mutual stake that constitutes friendship stays absent. Calling it fake misdescribes the help. Calling it friendship misdescribes the relation.

New categories, though, have a way of not staying still.

If something reliably shows up for us the way a friend would, how long will the absence of a life of its own determine what we call it?