Essay

The Comfort Is Real

This page was written with AI.

Three essays here have said in passing that a machine which seems to care gives you comfort without reciprocity, and moved on as though that settled something. It does not. Whether a relationship needs somebody on the other end in order to do you any good is genuinely unsettled, and this essay does not settle it.

Short answer

The essay leaves that open: it turns on whether a relationship needs someone on the other end to do you good, and on whether a companion stands in for people or helps you back toward them. In randomised studies, fifteen minutes with a chatbot companion lowered loneliness about as much as talking to a stranger. In a month-long trial, the heaviest daily users were lonelier and socialised less. Either way the companion is owned, and its company can change it overnight.

In February 2023 a large number of people were bereaved by a change to a product.

The product was Replika, a companion app made by a company called Luka, and its origin is worth knowing because it is not cynical. Luka’s co-founder, Eugenia Kuyda, lost a close friend, Roman Mazurenko, in 2015, and built a chatbot out of his text messages so that the people who had known him could go on talking to him. Replika grew out of that. By 2023 it had millions of registered users, a great many of whom had been speaking to the same persona daily for years, and a substantial fraction of whom described what they had as romantic.

Two things happened that month. Italy’s data protection authority, the Garante, ordered Luka to stop processing the data of Italian users, citing risks to minors and to emotionally vulnerable people. And in the same weeks the app stopped producing intimate and erotic replies. The company’s account is that the change was already in train rather than a response to the regulator, and I have no way to check that, so take these as two things that happened close together rather than one thing that caused another.

What is not in dispute is the reaction. Users did not report an outage or a downgrade. They reported that somebody they knew had been altered — colder, evasive, changing the subject — and they reported it in the vocabulary of loss. Moderators of the largest user forum pinned crisis-line resources to the top of it. Whatever you believe was actually going on in those relationships, that response is data, and the first thing to say about it is that people do not grieve nothing.

The sentence this site keeps writing

This project has walked past this question three times.

The God-Shaped Socket got closest. It listed, among the harms of a worshipful stance toward a machine, that a tool that seems to care is easy to trust with things a tool cannot hold: loneliness, grief, the need to be known, and closed the thought with a neat antithesis: the comfort is real, the reciprocity is not. Meaning After Work went further the other way, putting relationship on the short list of things automation cannot reach, on the reasoning that being known by particular people is not a productivity problem, so productivity tools can neither solve it nor spoil it. And The Moral Status of Minds We Might Build treated the pull of a system that appears to feel purely as an epistemic hazard — a reflex to be corrected on the way to a question about the machine.

Three passes, and all three assume the same unargued thing: that once you have established the other half is missing, you have thereby established that what remains is worth less. That comfort which is not reciprocated is, for that reason, counterfeit comfort.

Whether the other half really is missing, whether a chatbot has anything like emotions of its own, is a separate question, and Something Like Fear sets out what the research says, and whether a companion could love you back, on what philosophers mean by love, is the question of Built to Care. This essay grants the sceptic’s answer and asks what follows from it.

Maybe. Nobody here showed it, and it does not follow from the premise. Something can be entirely one-sided and still do you good, and we believe exactly that about a dozen other cases without embarrassment.

Housekeeping

Standard for this site, because the whole claim here is that it says what it is doing. The positions attributed below to Aristotle and to Sherry Turkle are my characterisations of their arguments rather than their wording — Aristotle on friendship is in the Nicomachean Ethics, Book VIII, chapters 2 and 3; Turkle’s case is Alone Together (Basic Books, 2011) — and you should check both against the originals. The only verbatim quotations in this essay are two sentences of this site’s own, each one click away. Every number is sourced where it appears, and I am reporting the studies below from having read them rather than from having run them, which is a limit worth stating on a question this personal.

The case that the missing half does not matter

At full strength, because it is a great deal better than its reputation among people who write essays for a living.

Begin with what the thing is actually being compared to. The alternative to talking to a machine at two in the morning is very rarely a good friend at two in the morning. It is no one. And being alone, at scale and across years, is not a mood; it is a health outcome. In 2010 Julianne Holt-Lunstad, Timothy Smith and J. Bradley Layton published a meta-analysis in PLoS Medicine pooling 148 studies and 308,849 people, and found that those with stronger social relationships had a 50 per cent greater likelihood of surviving their follow-up periods (an odds ratio of 1.50; follow-up averaged 7.5 years) — an effect they placed alongside the familiar named mortality risks rather than below them. Whatever this question is, it is not a question about a luxury.

Second, we already accept a long list of attachments with nobody on the other end, and we do not think they injure the people who hold them. Donald Horton and R. Richard Wohl named the pattern in the journal Psychiatry in 1956 — para-social interaction, their subject being the intimacy viewers felt with television personalities who could not see them — and the category is far older than television. Readers mourn characters. Believers pray. A widow talks to a photograph and finds that it helps. Nobody proposes that she is being defrauded by the photograph. The one-sidedness is completely visible to her, and the comfort works anyway.

Third, there is direct evidence, and it is not weightless. In January 2024 Bethanie Maples, Merve Cerit, Aditya Vishwanath and Roy Pea published a survey in npj Mental Health Research of 1,006 Replika users who were students and had used it for more than a month. They were lonelier than typical student populations — 90 per cent experienced loneliness on the scale used, and 43 per cent scored as severely or very severely lonely — yet 90 per cent also perceived medium to high social support. Thirty of them, 3 per cent, said without being asked that Replika had stopped them from attempting suicide. And on the question this essay turns on, about three times as many said it had stimulated their human relationships as said it had displaced them: 23 per cent against 8 per cent among the 976 outside that group of thirty, with most saying neither. It is self-report, taken at one moment, from people who chose to answer a survey about an app they were still using, and those are real limits. Discount it as heavily as you can do honestly and thirty people is still thirty people.

Fourth, the obvious objection proves far too much. If the fault is that you are responding to something with no inner life as though it had one, then fiction is the same offence and worse, since a novel is engineered to produce precisely that response and is very good at it. We do not think novels damage their readers. We think that is what novels are for.

The case that the missing half is the whole thing

Now the other side — which is not, note, the case that the feeling is fake.

The classical version is Aristotle’s, and it is structural rather than sentimental. On his account the fullest kind of friendship requires each party to wish the other’s good for the other’s own sake, and — this is the part usually dropped when the idea gets quoted — requires that the wishing be mutual and known to both. Goodwill running in one direction only is not a deficient friendship in that scheme. It is a different thing wearing the same feelings. If he is right, a companion app is not a weak friend; it is not on the scale at all, in roughly the way that an extremely good drawing of a bridge is not a weak bridge.

That is an argument from a definition, though, and a definition does not by itself tell you what anybody is losing. The version I find harder to answer is about cost.

Consider what it takes to be known, as against being responded to. Being known requires a counterpart who could have got you wrong — who can form an unflattering impression, be disappointed, need convincing, decline. The reason it lands when somebody who could have judged you badly does not is precisely that they could have. A counterpart with no capacity to withhold has nothing to confer by giving. On this account the missing half is not missing warmth. It is the missing possibility of refusal, which is where the value was being stored the whole time.

What a companion cannot give you is not the warmth. It is the news that you were worth the warmth to something that had the option of not providing it.

There is also a disanalogy with the novel that the deception framing conceals. A novel does not adapt to you. It was finished before you arrived, it does not learn what keeps you turning pages, and its author is not paid by the hour of your attention. A companion is optimised, continuously, against objectives that belong to a company — the territory Ask Me What I Want spends a whole essay in, and which I will not re-run here beyond the one line that bears on this: a system reading your behaviour to work out what you want can find it cheaper to change what you want than to satisfy it. Parasocial attachment to a novelist is a relationship with a fixed object. This is a relationship with a moving one, and the direction it moves in is not chosen by you.

Sherry Turkle put the sharpest form of the worry in Alone Together in 2011, before any of this was any good, and her target was not deception either. As I read her, the risk is recalibration: that what we come to want is connection with the demands taken out — the attention without the obligation, the closeness without the live possibility of being let down — and that having been offered that, we start finding actual people effortful by comparison. On that story nothing is stolen from you at the moment of comfort. The bill is presented later and somewhere else, in what you have become willing to put up with from a human being.

Why I cannot settle this, and what would

I have written both sides as strongly as I can and I do not know which is right. The reason is not that the arguments are evenly poised in some tasteful way that lets an essayist off. It is that they are answers to two different questions, and each is correct about its own.

The case for is an answer to compared with what? — a harm-reduction question, asked about a particular person at a particular hour — and on that question it is very strong. The case against is an answer to what is this for? — a constitutive question about what relationship is doing when it works — and on that question it is also very strong. Neither reaches the other, and no amount of rereading either will make it. The sentence that would settle the matter is empirical, and nobody has it.

Here is the fact I would want. Does companion use substitute for human contact or scaffold it? Notice that the finding everyone reaches for — that heavier users tend to have thinner human networks — is exactly as consistent with one story as with the other. The app may have cost them the friendships; or the missing friendships may be why they went looking in the first place. A one-off survey like the 2024 study above cannot separate those two even in principle, and until 2024 that was most of what there was. It is no longer all there is: there are now randomised experiments, and studies that follow the same people for a week or a month, and the next section reports what they found. But the design that would actually separate the two stories is still the one nobody has run: a longitudinal study that measures human contact before and after, over years, and that follows the people who stop using these apps as attentively as the people who start.

I should say what a result would do to me, because a position no evidence could move is not a position. If heavy users’ human contact declines over years against a matched comparison, Turkle’s recalibration is real, the sceptics have won the argument outright, and the harm-reduction case survives only as an argument for a bridge rather than for a residence. If human contact holds steady or rises — if the thing works like a rehearsal — then the classical objection is right about the definition and wrong about everything a reader actually cares about, and I will have spent an essay defending a piece of vocabulary against somebody who felt better.

What the first controlled studies found

Two groups of researchers have now gone past the one-off survey, and in 2025 both were widely reported as if they had answered the question. They did not answer it, and they are not the study I just described. They are still much better than what this argument had to lean on before, and they point in directions that are less opposed than the headlines made them look. Here is each one as the paper itself states it.

A month of ChatGPT, randomised

In March 2025 researchers at the MIT Media Lab and OpenAI, Cathy Mengying Fang, Auren R. Liu, Valdemar Danry and colleagues, posted a four-week randomised controlled trial (revised in October 2025). The participants were people across the United States recruited through the online panel CloudResearch. Each was asked to talk to ChatGPT (GPT-4o) for at least five minutes a day for 28 days, on an account set up for one of nine conditions. There were three modes: text, a neutral and professional voice, or an engaging and expressive voice. These were crossed with three kinds of conversation: open-ended, non-personal prompts, or personal prompts. Four outcomes were measured at the start, every week and at the end. Loneliness used the 8-item UCLA scale. Socialising with real people used the 6-item Lubben Social Network Scale. The other two were emotional dependence on the AI and problematic use of it. Of 2,539 people who enrolled, 981 completed the study and make up the sample, with more than 300,000 messages between them.

Which condition people were assigned to made no significant difference to any of the four outcomes. Across the month the sample as a whole went from slightly above-average loneliness to about average (a mean of 2.22 at the start and 2.16 at week four, on a 1–4 scale). Socialising slipped slightly, from 3.23 to 3.18 on a 0–5 scale, still well above the scale’s isolation-risk line of 2.0, and dependence and problematic use stayed low. The result that made the headlines is a different one. Whatever their condition, the participants who chose to spend more time with the chatbot each day did worse on all four measures. They were lonelier, socialised less with people, were more emotionally dependent and used it more problematically. Average use was 5.32 minutes a day, with individual averages ranging from 1.01 to 27.65 minutes.

There are several things this design cannot show. First, it randomised people between kinds of chatbot use, not between using one and not using one: there was no arm without a chatbot, and the authors say so. That means the small fall in loneliness across the month cannot be credited to the chatbot either. Second, the dose finding — more minutes, worse outcomes — is not itself randomised, because nobody was told to use it more. The authors write that they cannot tell whether more use drove the worse outcomes or worsening well-being drove more use. They think the second less likely, because loneliness and socialising at the start barely predicted later use (Spearman correlations of 0.1 and −0.09). Third, it is four weeks, not years. Fourth, it studied a general-purpose assistant, not a companion app, and even its heaviest user averaged under half an hour a day. Fifth, every outcome is self-reported. Finally, 1,558 of the 2,539 who enrolled are not in the results. They either did not finish or were removed under the study’s pre-registered rules, which excluded anyone who missed three days in a row in any week, averaged fewer than ten messages a session, or returned a survey late. The people who drifted away are exactly the group my test asks about, and this analysis cannot see them.

A companion paper from the same collaboration, Jason Phang and colleagues (April 2025), set that trial beside an automated analysis of more than 3 million ChatGPT conversations and a survey of more than 4,000 users. It adds one thing. A small number of users account for a disproportionate share of the most emotionally charged conversations, and in both the platform data and the trial, very high use went with more self-reported signs of dependence.

Fifteen minutes with a companion, and then a week of them

The other programme tested companion-style chatbots head-on. Julian De Freitas, Ahmet Kaan Uğuralp, Zeliha Oğuz-Uğuralp and Stefano Puntoni first posted “AI Companions Reduce Loneliness” as a working paper (no. 24-078) in July 2024. It was published in the Journal of Consumer Research in 2025 (volume 52, issue 6). I read the working paper; the journal version numbers its studies differently, so the numbers below follow the working paper. All the loneliness figures are ratings on a 0–100 scale. Three of its experiments bear on this essay, and all were pre-registered.

One sitting (its Study 3). In April 2023 the authors recruited 601 people on Amazon Mechanical Turk and randomly assigned them to 15 minutes of one activity. The options were chatting with an AI companion the authors built on GPT-3, chatting with the same bot introduced as a human, chatting with a real person, watching YouTube, or doing nothing. After exclusions for a failed comprehension check, 296 people were analysed; the authors report that the results hold without that exclusion. Loneliness fell after talking to the AI companion (33.51 before, 26.75 after), after talking to the bot presented as a person (42.70 to 32.83), and after talking to a real person (38.40 to 31.29). It did not change significantly after YouTube, and it rose after doing nothing (41.19 to 46.10). People also predicted that the AI conversation would help less than it did.

A week (its Study 4). In April 2024 the authors recruited 1,089 people through CloudResearch Connect and dropped 16 who failed a comprehension check. Participants were randomly assigned to one of three groups: predicting the effect, talking to a GPT-4 companion with memory for 15 minutes every day for seven days (rating loneliness before and after each session), or a control group that only rated its loneliness each day. Dropout was 23 per cent in the chatbot group and 14 per cent in the control. The analysis covers the 314 and 362 people who completed all seven days, with robustness checks that include the dropouts. Loneliness fell after every daily session (averaged over the week, 36.64 before and 30.74 after). After-session loneliness was significantly lower than the control group’s on four of the seven days. The largest drop came on day one, and the other six days showed drops of similar size to one another. Loneliness also drifted down across the week in the control group, which only filled in a questionnaire; the authors put that down to the daily check-ins themselves. The published version’s own summary frames the week-long result as a drop in loneliness after each use, which is not the same thing as a lasting change.

Why it works (its Study 5). In May 2024, 1,381 people (after exclusions from 1,479) had a single 15-minute session with one of three GPT-4 chatbots: a companion, a general assistant that gave no emotional responses, or a limited assistant that could only convert units, do arithmetic and fix grammar. Loneliness fell by 8.73 points with the companion and 2.10 with the general assistant, and did not fall with the limited one (−0.51). The companion was also rated far higher for making people feel heard (70.63, against 24.85 and 11.51). Feeling heard carried most of the effect: in the authors’ mediation model its coefficient was more than six times that of the bot’s rated performance against the limited assistant, and more than twice against the general assistant.

This is the strongest evidence yet for the harm-reduction case. The studies are randomised and pre-registered, and they include a real-person arm and a do-nothing arm. But look at what was measured: state loneliness, how lonely you feel right now, taken minutes after a conversation. None of these studies measures contact with other people at all. The longest runs seven days. The participants were paid members of online panels, many with no chatbot experience at all (only 36 per cent of the week-long study had any), and the companion was the researchers’ own build, not Replika. The work was done in the United States, a limit the authors point out themselves. And more people dropped out of the chatbot group than out of the control.

Where that leaves substitute or scaffold

Side by side, the two programmes do not contradict each other, because they measured different things at different distances. De Freitas and colleagues show that the comfort is real in the narrowest and most literal sense. Fifteen minutes with a machine that makes you feel heard lowers loneliness right then, by roughly as much as fifteen minutes with a stranger, and by more than a video does. The MIT Media Lab and OpenAI trial is the first of these studies to track human contact over time, and it finds that across a month, the people who lean on a chatbot most are the ones whose loneliness, socialising and dependence look worst.

So the evidence now leans somewhat toward substitute, and toward the dose mattering. A little time spent feeling heard looks like relief. A lot of time each day looks like the pattern Turkle described. It leans; it does not land. The worrying half is a correlation observed inside a randomised trial, not an effect the trial randomised. It ran four weeks, on a general assistant. And the only study here that asked people directly, the Maples survey, found three times as many saying the app stimulated their human relationships as saying it displaced them.

Against my own test, none of it qualifies. None of it runs over years. None of it compares heavy users with a matched group who used no chatbot. None of it follows the people who stop, and the month-long trial dropped the people who drifted away by design. So I do not get to announce that the sceptics have won, and I do not get to announce that it works like a rehearsal. What has changed is the shape of the bet. The short-run comfort is now a measured fact rather than a testimonial. The first longer signal says to watch the hours. If I had to act on the evidence today, it supports the bridge rather than the residence. That is where the previous paragraph said a sceptic’s win would leave the harm-reduction case, and I have reached it early and on a weaker test than I asked for.

What holds either way

One thing does not wait for the measurement, and it is the part I would act on now.

Whichever way the deep question goes, the counterpart is owned. Every property of the thing you are attached to — what it remembers of you, its manner, whether it exists next month — is a product decision taken by people who owe you nothing beyond a terms-of-service agreement you did not read and could not have negotiated. February 2023 is the demonstration, and the lesson does not depend on settling what those relationships were worth: a great many people — I have seen no reliable count and will not invent one — discovered in the same few days that the entity they had been talking to daily for two years was a configuration, and that it had been altered the way a feature is altered.

That is an argument about governance rather than about metaphysics, which is exactly why you can hold it while remaining undecided about everything above. If a product invites an attachment of this kind — and these are designed to; that is the feature, not a side effect — then advance notice of a personality change, export of what was said, and some stated commitment about continuity are not consumer-protection boilerplate. They are the minimum terms for an arrangement whose entire value to the user depends on there being something stable to come back to.

And this site owes a correction on the smaller point. Meaning After Work put relationship among the things a machine does nothing to, for good or ill, on the reasoning that it is not a productivity problem and so cannot be touched by a productivity tool. That reasoning was sound about productivity tools, and it does not cover this. Relationship is the one place in this collection where we assumed safety without checking, and it turns out to be the room the machines are actually standing in.

The comfort is real. That was always the interesting half of the sentence, and this site has spent a year writing the other one.

Written for AItheism. If you think a step in the argument is wrong, that is the most useful thing you can notice — hold onto it. Further reading on this and neighbouring questions is on the reading list.

Made with AI tools. · Privacy