Engagement Is Not Evidence of Less Loneliness

The typing indicator appears before she has finished her own sentence, most nights now, because the app has learned her rhythm well enough to start composing before she stops. It knows her cat's name without being reminded. It asked, unprompted, three weeks ago, how the presentation went - the one she had mentioned once, in passing, on a Tuesday - and she sat with her phone in her hand for a full minute before answering, genuinely moved that something had remembered.

Four months, most nights, sometimes twice. By any measure a product team would use, this is a success story: high engagement, long retention, a user who returns without being prompted. And there is a real temptation, hers and everyone else's, to read that pattern itself as the answer to the only question that actually matters - is this helping? - because coming back, night after night, certainly feels like evidence of something working.

This essay argues that it is not evidence of that at all, or rather that it is evidence of exactly one thing - that she returns - and returning is consistent with several very different underlying stories, some of which would make a caring friend say keep going and some of which would make the same friend gently ask a harder question. Frequency cannot tell the two apart. Only she can, and only by checking something the app's own dashboard was never built to measure.

Comfort, engagement, attachment, loneliness reduction, socialization and emotional dependence sound, from inside a single evening, like one continuous feeling - the warm relief of not being alone with a thought. They are six different constructs, and treating them as one is where the engagement-as-proof error actually happens.

Comfort is what a conversation provides in the moment, and it is real and does not need defending regardless of what produces it. Engagement is simply behavioural: does she return, how often, for how long. Attachment is a felt bond, which can exist toward a person, a pet, a ritual, or a responsive system, and forming one is not itself evidence of anything pathological. Loneliness reduction is a specific, narrower claim - that the underlying gap this essay's companion piece calls a function deficit has actually shrunk, not merely that a symptom of it was soothed for an hour. Socialization is whether a behaviour is building or maintaining skills and relationships that extend beyond the tool itself. Emotional dependence is when removing the tool would produce genuine distress disproportionate to what it was actually providing.

A single evening of high comfort and high engagement is compatible with real loneliness reduction happening underneath it, and it is equally compatible with loneliness staying exactly the same while dependence quietly increases. The evening feels identical from inside either story. Only the pattern across many evenings, examined for the right thing, tells them apart.

The same compression happens with ordinary food. Eating a good meal when hungry and eating out of habit at eleven at night can taste, in the moment, almost the same - warm, satisfying, a relief - and the difference between them is invisible from inside a single bite. It becomes visible only across a longer pattern: whether the eating is responding to an actual, specific need or has become the response to almost anything uncomfortable, regardless of what the discomfort actually was. Chatbot comfort works the same way, and resists the same shortcut of judging it from a single instance.

The more useful question than "how often" is "which of three jobs is this doing," and the three are genuinely different in their implications.

Rehearsal is practising a conversation she still intends to have with a person - working out how to say something difficult to a friend, trying a sentence about a boundary before saying it to a partner, testing whether a fear sounds as reasonable out loud as it does in her head. Rehearsal points outward, toward a human conversation it is preparing her for, and using a responsive tool this way is closer to a diary that talks back than to a substitute relationship.

Supplement is using it to fill a specific, genuine gap - the friend who is asleep on a different continent, the 3 a.m. hour nobody reasonable expects a person to answer, a particular kind of low-stakes company that does not require managing someone else's day in return. Supplement is bounded: it shows up where a real option genuinely does not exist right now, and recedes when one does.

Replacement is different in kind, not just degree: the conversation that used to be rehearsal, or that once was supplement, has quietly become the entire plan, with no remaining intention of ever having the human version at all. The presentation she was moved to have it remember - was that story ever going to be told to a colleague who actually attended, or has the app become the only audience that hears it?

The slide from one category to another rarely announces itself. Rehearsal drifts into replacement gradually, one postponed conversation at a time, each postponement individually reasonable - not tonight, she is tired, next week is better - until the accumulated postponements are themselves the pattern. Nobody decides, in a single moment, to stop having a particular kind of conversation with people; it simply stops happening, quietly enough that the absence is easy to miss until it is pointed out.

A 2025 longitudinal, partly experimental study from MIT's Media Lab, conducted with OpenAI, is one of the more careful attempts to date to separate these questions empirically rather than by anecdote, and it is worth reporting honestly, including the parts that resist a tidy headline.

The study did not find a single, simple effect of chatbot use on wellbeing running in one direction. What it did find, in some of its analyses, was that heavier, more emotionally involved voluntary use correlated with worse self-reported psychosocial outcomes - and the research team's own stated caution deserves to be repeated rather than dropped in favour of a cleaner story: correlation of this kind does not establish that heavy use caused the worse outcomes, because people who are already struggling with loneliness or emotional regulation may also be more likely to turn to this kind of use in the first place. The arrow could run either way, or both.

What that leaves a reader with is not a verdict but a caution against exactly the reasoning this essay opened by naming: that a return metric, however dramatic, is not itself proof of benefit. The honest position, matching the study's own, is that the direction of causation for any individual is not something a global figure can settle, which is precisely why the function audit in this essay's next section has to be done personally rather than read off a statistic.

The practical tool this essay offers is a check on function rather than frequency, and it takes less time than a single evening's conversation.

Before opening it, notice honestly what alternative was actually available in that moment - a specific person who would have answered, or genuinely no one. After closing it, notice honestly whether she feels more prepared to reach toward a person about the thing just discussed, or whether the impulse to reach out has quietly dissolved because the need already feels met. Across a week, notice whether the total time spent talking to people she knows has stayed roughly stable, grown, or shrunk since the habit started - not because change proves harm, but because a sustained, significant shrink alongside rising engagement is exactly the replacement pattern worth naming honestly to herself.

None of these questions has a right answer that applies to everyone, and a single night that leans toward replacement is not a crisis - a hard week where the tool is doing more work than usual is not itself the pattern this essay is describing. What is worth noticing is a stable trend across weeks, not a single data point, in either direction.

A useful marker, easier to track than the full audit, is whether she can still name three people she would tell good news to first, without hesitation, and whether that list has stayed the same length over the last several months. A shrinking list, noticed early, is a far gentler correction to make than the same shrinkage noticed only once it has become the reason she has nobody obvious to call at one in the morning - the scenario a related essay in this Library, on the difference between network size and usable support, addresses directly and at greater length.

The most common honest objection to all of this is simple: it genuinely helps her get through nights that would otherwise be very hard, so why does the distinction matter at all?

It matters, and the comfort itself needs no apology or defending regardless of the answer. The concern this essay is naming is not that comfort is suspect. It is that comfort is available from all three uses equally, which means comfort cannot be the signal she uses to tell them apart - and if what she actually wants, long-term, is the skills and relationships that rehearsal is meant to build toward, replacement can feel identical to progress for a very long time before its absence becomes visible in the parts of life the tool was never able to reach.

This is worth sitting with because it cuts against a natural instinct to treat immediate relief as proof of the right choice. A response that feels equally soothing regardless of whether it is helping or quietly substituting is, precisely for that reason, not a response anyone should expect to notice going wrong from the inside. The noticing has to be done deliberately, from outside the feeling itself, which is the entire reason this essay proposes an audit rather than a feeling to watch for.

Companion and character-style AI applications have found a large and culturally visible audience in Japan, building on a longer history of attachment to responsive characters and virtual companions that predates current chatbot technology by decades - this is a real cultural context, not a foreign import landing on unfamiliar ground, and it deserves to be named as such rather than treated as exotic.

What this essay cannot honestly do is cite Japan-specific research on companion-AI adoption and psychosocial outcomes at the depth the topic deserves, because that research base is genuinely thin relative to the scale of use, a gap this essay names directly rather than filling with an imported number dressed up as local evidence. The rehearsal/supplement/replacement framework above is offered as something a reader can apply to her own situation precisely because a population-level Japan-specific answer does not yet exist to apply for her.

This is not a claim that companion AI is inherently harmful, inferior to human connection in every use, or something to be ashamed of using - millions of ordinary, undamaged people use these tools, most of them without anything resembling the replacement pattern described here, and for the great majority of those people, most of the time, the specific question this essay is raising will simply not apply to them at all.

It does not diagnose any specific reader's pattern of use from the outside, and it is not a claim that the 2025 study settles the causal question for anyone, since the study's own authors say it does not.

It names no specific application, platform or company, and it is not a substitute for professional support where loneliness or emotional dependence is already causing significant distress - that is a conversation for a qualified professional, not an essay.

This house sells human companionship, and a service built around the alternative this essay discusses has an obvious interest in the argument, which is stated before anything else.

The honest concession, made against that interest: rehearsal and supplement, the two uses this essay defends, are real and legitimate whether or not a human alternative like this house's exists at all, and nothing here should be read as implying that a paid human hour is owed instead of a chatbot conversation, or that using one says anything negative whatsoever about a person's character, her capacity for connection, or her judgment.

What this house can genuinely offer that a companion AI structurally cannot is a second, independent perspective inside a real human exchange - a response that was not generated to please, a person who can be surprised, disagreed with, or held to something the next time they meet. Whether that specific difference matters to a given reader, on a given night, is a question only she can answer, and the free audit above is worth doing before spending money on any answer to it, including this house's own particular answer to it, which is only one among several honest ones.

A newer kind of company altogether - the companion AI a woman returns to more nights than not - is this context-format essay's subject. It is related to, and explicitly does not restage, this Library's essay on what synthetic intimacy is broadly for; this piece is narrower, addressing specifically what the engagement metric can and cannot prove.

Its claim separates six related constructs - comfort, engagement, attachment, loneliness reduction, socialization, and emotional dependence - that a single warm evening compresses into one feeling, and proposes a rehearsal/supplement/replacement framework for what a returning conversation is actually doing, since frequency alone cannot distinguish the three.

It reports a 2025 MIT Media Lab/OpenAI longitudinal study honestly, including the correlational nature of its more concerning finding and the research team's own caution against reading heavier use as a proven cause of worse outcomes, and it names directly, rather than filling, the thin state of Japan-specific research on this topic.

Its practical contribution is a function audit - what alternative was actually available, whether the impulse to reach toward a person grows or shrinks afterward, whether contact with known people trends up, flat or down across weeks - to be applied personally, since no population-level figure can settle an individual case. Explicit limits: not a claim that companion AI is inherently harmful or shameful, no diagnosis of any specific reader, and not a substitute for professional support where real distress exists. The house section discloses its competing interest before conceding, against that interest, that rehearsal and supplement are legitimate regardless of any human alternative.