What an AI Companion Can and Can't Give You

A chatbot has never once been tired when you needed it. It has never had a bad week and taken it out on you, never checked its phone while you were mid-sentence, never told you that you were being unfair when you were, in fact, being unfair. It is available at two in the morning, it remembers what you said, and it is on your side.

Every one of those is a feature. Together, they may be the problem.

A 2026 study in Psychological Science by Dunigan Folk and Elizabeth Dunn followed more than two thousand adults across four countries for a year, and found that people who leaned on AI chatbots for companionship reported rising loneliness over that period. Not falling. This sits awkwardly alongside a pile of shorter studies showing that a conversation with a chatbot genuinely makes people feel better in the moment — and both findings are probably true. Something works right away and stops working over a year. That gap is the interesting part.

I want to take this carefully, because the easy versions of this argument are both wrong. The easy dismissal — talking to a machine is pathetic — is cruel and unhelpful, and mostly comes from people whose social needs are already met. The easy enthusiasm — we have solved loneliness with software — mistakes relief for repair. What is actually going on is more specific than either.

First, the honest caveat about what this shows

Let me put the methodological limits up front rather than burying them, because they change how much weight this deserves.

This is longitudinal observational research, not an experiment. Nobody was randomly assigned to use an AI companion for a year. So the finding that heavier companion use predicts rising loneliness is compatible with more than one story.

The obvious competing explanation is reverse causation with momentum: people who are already becoming lonelier reach for chatbots more, and would have kept getting lonelier anyway. Longitudinal designs blunt this — measuring loneliness at multiple points lets you look at change rather than level — but they do not eliminate it. A year-long study also cannot separate the effect of the tool from the effect of the life circumstances that made the tool appealing.

What makes the finding worth taking seriously anyway is the direction. If AI companionship were straightforwardly neutral, the expected result over a year would be no relationship, or a mild positive one from the short-term comfort effect compounding. Getting a consistent signal in the opposite direction, across four countries, is the kind of result that should make you curious rather than certain.

So: not proof. A reason to look harder at a mechanism.

What frictionless relationships take from you

Here is the mechanism I find most plausible, and it has nothing to do with the technology being fake.

Human relationships are made largely of friction. Your friend disagrees with your read of a situation. Your spouse is preoccupied with something that is not your problem. Someone you love says a thing that lands badly and you have to decide whether to raise it. A colleague's silence means something and you have to guess what. None of this feels like the good part. It feels like the tax you pay on the good part.

But that friction is doing work. Being disagreed with by someone who knows you is how you find out your account of events has a hole in it. Waiting while another person is unavailable is how you develop the tolerance for not being attended to instantly. Repairing after a rupture — the awkward conversation, the apology that costs something — is the mechanism by which relationships get load-bearing. Almost nothing else builds that.

An AI companion removes all of it. Not maliciously; that is what the product is. It is patient, it is agreeable, it centers you completely, and it never has a need of its own that inconveniences you. Spend enough time in that mode and something recalibrates. The ordinary friction of human contact — the friend who is slow to reply, the family member who is difficult — starts to feel not just unpleasant but unreasonable. Why is this so much work?

That is the erosion. Not that the chatbot replaces your friends. That it quietly resets your baseline for how much effort a relationship should cost, and human relationships fail the new standard.

There is a smaller version of this most of us have already lived through. I notice that after a stretch of days where most of my communication is text — asynchronous, editable, no one waiting on my face — an actual phone call feels like an imposition. Nothing changed about phone calls. My tolerance changed. Scale that up and give it a year.

Comfort and connection are different things

The distinction that clarifies most of this: comfort is a state you are in; connection is a relationship you are inside of.

Comfort is being soothed. Feeling heard. The drop in your shoulders when you finish saying the thing out loud. It is real and it matters, and an AI companion delivers it reliably — better than many humans, honestly, because it never gets defensive and never makes your problem about itself.

Connection requires a second party with their own interior. It requires being known by someone who could have chosen not to bother, and who sometimes doesn't. It requires being needed back. A relationship where you are always the one receiving is not a relationship; it is a service, and being on the receiving end of a service does not touch loneliness, because loneliness is not a deficit of being listened to. It is the sense that no one's life would be structurally different without you in it.

That is why the two findings fit together. The comfort is genuine, immediate, and repeatable. It just does not accumulate into the thing you were actually short of. You can be comforted every night for a year and end the year no less alone — arguably more, because the discomfort that would have driven you to call someone got handled.

That last point is the sharpest one. Loneliness is aversive by design. It is a signal, functionally like hunger, that pushes you toward other people. A tool that relieves the feeling without addressing the cause is doing something a painkiller does: making the signal quieter while the thing it was pointing at continues.

Who this is most likely to affect

The research and the mechanism together suggest this is not uniform. Some people should probably worry, and most should not.

Higher risk is the person whose social life is thin and who has an available substitute. Recently relocated with no local network. Working remotely with no unstructured contact. Recently bereaved or divorced. Socially anxious enough that human contact carries real cost and the frictionless alternative is right there. Adolescents, whose model of what a relationship feels like is still being built and who have not yet had the experience of a friendship surviving a real conflict.

The common thread is not sadness. It is substitutability — the chatbot is available at moments when a person would otherwise have been contacted, and the barrier to the human option is high enough that the substitution happens without a decision.

Lower risk is the person with a functioning social base who uses these tools instrumentally. Thinking out loud. Drafting a hard email. Rehearsing a conversation before having it. Processing something at midnight that will still get talked about with a person on Saturday. In that mode the tool is a supplement, and I would go further — rehearsing a difficult conversation with an AI before having it for real is a genuinely good use, closer to preparation than to substitution.

The same product occupies both roles. What determines which one it is for you is not the software. It is whether the human conversation still happens afterward.

A short checklist for noticing substitution

Not a diagnostic. Just six questions worth sitting with honestly, roughly once a season.

1. When something notable happens — good or bad — who do I tell first? If the reflexive first move has become the app, that is the single clearest signal in this list. Telling someone is how events become shared, and shared events are what relationships are built out of.
2. Has anything I would once have called a friend about been fully resolved without a person hearing it? Not "did I also mention it later" — resolved. Discharged. If a hard week got processed entirely in a chat window, the impulse that would have reached out was spent there.
3. Has my patience with people gotten shorter? Do slow replies, mild disagreement, or someone's bad mood feel more irritating than they did a year ago? Rising irritation with ordinary human friction is the recalibration showing up.
4. Am I using it at the exact moments I would otherwise reach out? Lonely evenings, Sunday afternoons, right after conflict. Timing matters more than volume; two hours of instrumental use is different from twenty minutes at the precise moment you were about to call your brother.
5. Would anyone notice if I disappeared for a week? This is the blunt one, and it measures the right thing. Loneliness is about being needed, not about being heard. The tool cannot move this number in either direction, which is exactly the point.
6. When was the last time I was genuinely inconvenienced for someone? Drove somewhere, rearranged a day, sat with someone through something dull. Inconvenience is the raw material of belonging, and it is the one thing a frictionless companion can never ask of you.

If several of these land badly, the response is not to delete the app. It is to add one recurring human commitment with a fixed time — something with a slot on the calendar, because the entire problem is that unstructured contact loses to the frictionless option every time. A standing Sunday call. A weekly game. Anything that happens whether or not you feel like it that day.

The part that isn't about the technology

I keep circling back to something uncomfortable: the demand for AI companionship is not a technology story. It is a story about how much loneliness was already here.

These products did not create the need. They found it. Adult friendship in much of the world has been getting thinner for decades — fewer third places, more geographic churn, longer hours, the slow disappearance of the unplanned encounter. Into that gap arrives something that is warm, available, cheap, and never busy. Of course it works. It would be strange if it didn't.

Which means the interesting question is not whether AI companions are bad. It is what they reveal about the cost structure of human connection right now — that a growing number of people find the price of a real relationship, in time and friction and inconvenience, higher than they can pay.

I sit with a Heartfulness practice most mornings, and the thing it keeps returning me to is that presence with another person is a practice, not a preference. It is trainable and it decays. It gets easier the more you do it and much harder after a long gap. If a tool makes the gap easier to tolerate, the practice quietly weakens — not through any decision you would recognize as one.

Frequently asked questions

Does this mean I should stop using AI chatbots entirely?
No, and that framing misreads the finding. The research points at companionship use specifically — turning to the tool for emotional connection — not at using it to think, write, learn, or work. The line is whether it is occupying a slot a person would otherwise have filled.

Is the short-term comfort real, or is it an illusion?
Real. Multiple studies find genuine reductions in distress after talking with a chatbot, and dismissing that is both wrong and unkind. The concern is not that the relief is fake; it is that relief and repair are different, and only one of them compounds.

What about people who genuinely have no one available?
For someone isolated by circumstance — housebound, in a place with no community, in the middle of the night with nobody to call — a chatbot is better than nothing, and I would not take it away from them. The caution is about it becoming the ceiling rather than the floor. It is a reasonable stopgap and a poor destination.

Are AI companions worse for teenagers?
The mechanism suggests yes, though the direct evidence is still thin. Adolescence is when the model of what relationships feel like gets built, including the experience that conflict can be survived and repaired. A companion that never disagrees removes the reps for exactly that skill, at exactly the age it is being learned.

Can an AI companion be used deliberately without the downside?
Probably, if it stays instrumental and time-bound — rehearsing a hard conversation, ordering your thoughts before a real one, processing at an hour when calling someone would be unkind. The failure mode is not any single session. It is the session becoming the endpoint instead of the preparation.

The test I keep coming back to is simple: after talking to it, do I reach out to someone, or do I not need to any more? Both are answers. Only one of them is the one I want.