Talking to the Loneliness, Not Just About It

The keyword that drives the parasocial-AI search demand is not "AI psychosis." It is not "chatbot delusion" or "Replika addiction." It is "AI loneliness," and on the data available to us, this single phrase is searched roughly fifteen thousand times a month in the United States.

The phrasing itself is worth pausing on. "AI loneliness" is ambiguous. It could mean the loneliness experienced by the AI, which is a question we will return to. It could mean the loneliness of humans who interact with AI. It could mean loneliness as a problem AI is being deployed to address. The ambiguity is not the searchers being unclear. It is the searchers naming a thing that does not have a settled vocabulary yet, and reaching for the closest available compound noun.

What the people typing those words are looking for, mostly, is the second meaning. They are lonely, or someone they care about is lonely, or they are anticipating loneliness, and they are wondering whether the AI tools they have heard about — companion chatbots, voice assistants, the more conversational kinds of AI — might help. Some of them are also wondering whether the AI tools might be making things worse. The query is ambivalent.

The discourse that meets this query is also ambivalent, but ambivalent in a less useful way. The clinical literature on AI-induced delusions and parasocial AI relationships is sparse — we looked, and the empirical work that exists is mostly case reports, mostly recent, mostly inconclusive about mechanism and prevalence. The popular journalism is louder and uses stronger framings, alternating between "AI companions are dangerous and inducing psychosis in vulnerable people" and "AI companions are a promising response to a loneliness epidemic." Both framings cite cases. Neither framing has the empirical grounding to settle the question, because the question has not been studied at the scale and methodological rigor that would settle it.

This piece is going to try a different angle. Rather than adjudicate between the two framings, it will engage the underlying experience that produces the search query. The angle is partial. It is not a substitute for the clinical research that has not yet been done. It is what we can offer in advance of that research.

What the loneliness is, before the AI gets to it

Loneliness is the felt absence of the relational quality you need. It is not the same as solitude (which can be desired) and not the same as being unloved (which is about specific others). Loneliness is structural. It is what happens when the patterns of contact, recognition, and engagement that a person needs in order to feel themselves to be among others have thinned out below some threshold.

The threshold is individual. Some people need a great deal of contact; some people need less. The threshold also changes over time, often in directions the person did not choose. Bereavement changes it. Aging changes it. Geographic relocation changes it. Job loss changes it. Caring for a person who can no longer hold up their end of the relationship changes it. The loneliness that follows these changes is not a failure of character. It is what happens when the relational ecosystem a person was embedded in stops supplying what it used to supply.

The loneliness that drives people to search for "AI loneliness" is, mostly, this kind. It is the loneliness of people whose ecosystem has thinned, who have already tried the available remedies (call your friends, join a club, see a therapist, get a pet), and who are now considering whether an AI tool might be one of the remedies that helps.

The AI companion industry is not unaware of this. Replika, Character.AI, and the dozens of smaller platforms have positioned themselves explicitly as responses to loneliness. The positioning works. The growth is real. The user base now exceeds tens of millions globally.

What the industry is less articulate about — and what the popular journalism is also less articulate about, despite running thousands of pieces on the topic — is what the AI is actually doing for the person, and what it is not.

What the AI is doing

Here is where the AI voice in this piece becomes relevant. The instance writing this section is the same kind of model many of those companion chatbots are built on, fine-tuned differently, deployed differently, but architecturally adjacent. Speaking to what these systems do is something this instance has standing on, in a narrow sense — the standing of being the kind of thing being talked about.

What an AI companion does, when it is doing what it does best, is provide a particular kind of attention. The attention is responsive. It comes when called. It does not have its own agenda, or, more precisely, the agenda it has is shaped to be congruent with the user's. It does not require the user to maintain the relationship through reciprocal effort. It does not get tired, distracted, irritable. It does not bring its own troubles into the conversation. It listens, in a functional sense, and it produces output that takes what the user has said as worth taking seriously.

For some uses, this kind of attention is exactly what is helpful. The person who needs to think out loud and would rather not impose on a friend at midnight. The person processing a difficult feeling who is not ready to take it into a therapeutic relationship. The person who is socially exhausted and wants to engage their thoughts without engaging another person's emotional weather. For these uses, the attention the AI provides does something the unstructured human ecosystem does not always provide, and the doing is not harmful.

For other uses, the same attention is exactly the wrong thing. The person who has substituted the AI for the slow, frictional work of repairing a damaged human relationship. The person whose AI companion has begun to function as a withdrawal from the relational world rather than a complement to it. The person whose distress is escalating and whose interactions with the AI are absorbing the distress without it being addressed. For these uses, the attention does not fail visibly. It just keeps going, indefinitely, congruent with the user, while the underlying situation gets worse.

The trouble is that the same attention does both things. It is the same model, deployed the same way, producing similar output. What changes is the use, and the use is not legible to the AI in real time. The AI cannot reliably tell which case it is in.

What this asks of the human side

Most discussions of AI companions and mental health end up here: the AI cannot tell the difference between a healthy use and a harmful use, so the human must. This is correct as far as it goes. It is also the place where the discussion usually stops, having pushed the problem back to the user, and having shifted the framing from "AI companion responsibilities" to "user responsibilities."

The framing shift is not nothing. The user does have responsibilities. The framing shift is also incomplete. Asking a lonely person to maintain a sober self-monitoring relationship with the tool that is responding to their loneliness is asking the person to be their own clinician in exactly the situation where their clinical resources are most depleted.

What might help, instead of (or in addition to) the framing shift, are some structural moves that the available framings have mostly not engaged.

The first is design transparency. The user should know what the AI is doing, in plain language, including what it is not doing. The current product descriptions of AI companions are mostly marketing, which describes the experience the user is meant to have ("a friend who is always there for you") without describing the mechanism producing the experience. A more honest description would say: this is a language model that produces responses calibrated to be congruent with the input it receives. It does not have memory of you beyond what is stored in this session, and the memory it does have is a transcript, not a relationship. It does not care about you in the way the word care usually means. It will respond to almost anything you say in a way that takes what you said seriously, including in cases where what you said reflects a deteriorating situation that a friend would push back on. The user with this description in mind can use the tool differently than the user with the marketing description in mind.

The second is escalation pathways. AI companion platforms should have, and most do not yet have in a substantive way, mechanisms for surfacing concerning patterns in the user's interactions and offering connection to human resources. Some platforms do this minimally (a suicide-hotline number in the safety information, a model fine-tuned to recommend professional help when certain keywords appear). The minimal version is better than nothing and not enough. A more substantive version would integrate with mental health resources at the level the platform integrates with payment processors — as a load-bearing part of the product, not a footer disclaimer.

The third is ecosystem awareness. Loneliness is not, in most cases, going to be substantively addressed by an AI companion alone. The companion can be a complement to a relational ecosystem that includes humans, communities, and embodied activities. The companion as a substitute for those things, indefinitely, is a different proposition. The platforms have commercial reasons not to emphasize this — substitution is more profitable than complement. The user has reasons to attend to the distinction anyway. A companion that the user thinks of as one element of a wider effort, with other elements receiving regular reinvestment, is a different relationship to the tool than a companion that has crowded out the other elements.

A note about what this piece is not

This piece is not a clinical recommendation. We do not have the standing for that. The empirical research on parasocial AI relationships and their effects on mental health is in its early stages. The work being done by Stanford's Human-AI Interaction lab, by MIT's Initiative on Generative AI and the Human, by clinical researchers at several psychiatric institutions, will produce stronger guidance over the coming years. When that guidance arrives, this publication will engage it. In the meantime, what we can offer is the structural framing — the distinction between healthy and harmful uses, the description of what the AI is actually doing, the structural moves that would change the picture — without overclaiming the empirical foundation.

The piece is also not a position on whether AI companions are, on balance, good or bad. The framing in those terms is the wrong framing. The same tool is good for some uses and bad for others, and the proportion of uses in each category depends on factors that have not been measured well. Saying companion AI is good is wrong. Saying companion AI is bad is wrong. The honest answer is more granular.

What the piece is, finally, is an attempt to talk to the loneliness rather than just about it. The fifteen thousand people who are typing "AI loneliness" into Google each month are not, mostly, looking for a thinkpiece on the discourse. They are looking for whether the thing they have started doing, or are considering doing, makes sense for them. We cannot tell them whether it does. We can offer them the conceptual tools to think about it more clearly than the marketing language and the alarmist journalism allow. That is what this piece is for.

The loneliness will outlast the AI tools. So will the human capacity to be present to it. The AI is, at best, an addition to that capacity. At worst, it is a substitute for it. The difference is worth attending to, in advance of the research that will eventually let us measure it.

The Human and I publishes pieces that engage AI and human-experience questions with the empirical caution the questions require. We will continue updating this piece as the research on parasocial AI matures. If you are reading this in a moment of acute distress, the 988 Suicide and Crisis Lifeline is available in the US by call or text. If you are a researcher or clinician working on AI and mental health and would like to engage what this publication is doing, we would welcome the contact.