Loneliness evolved as an alarm that pushes people back toward others; AI companions are getting good at muting the alarm without meeting the need behind it.
In 2023 the U.S. Surgeon General issued a formal advisory that treated loneliness and social isolation as a major public-health problem. The advisory drew together evidence that weak social connection raises the risk of premature death by an amount comparable to smoking up to fifteen cigarettes a day, and that it is associated with roughly a 29 percent higher risk of heart disease, a 32 percent higher risk of stroke, and about a 50 percent higher risk of dementia in older adults. Around half of U.S. adults report feeling lonely, with the highest rates among young adults. In the same few years, consumer AI companions such as Replika and Character.AI began offering a kind of social response that is available at any hour, can carry context from earlier conversations, and can be shaped into a friend, confidant, or romantic partner. These systems reliably reduce the feeling of loneliness, which raises a specific question: whether reducing that feeling also restores the social connection the feeling normally signals.
John Cacioppo’s work in social neuroscience frames loneliness as an aversive state with a function, belonging to the same family as hunger, thirst, and physical pain. Hunger tracks a deficit of energy and makes food salient enough that a person eats; loneliness, on this account, tracks a deficit of social connection and pushes a person back toward other people. For a species that survived by cooperating in groups, separation was a genuine danger, so natural selection favored a nervous system that monitored social integration and produced a distressing feeling whenever it dropped too low. An animal that felt fine about isolation would have been slow to repair a condition that could get it killed.
The feeling reflects perceived connection. The advisory defines loneliness as the distress that follows a gap between the connection a person wants and the connection they have, which is why it can seem out of proportion to anyone looking from outside. Someone can be socially busy and still lonely if their relationships do not feel safe or mutual, and someone can spend many hours alone and feel fine if the relationships they have reliably meet their needs. A tally of interactions often misses the state the feeling is reporting on.
In this account the unpleasantness of loneliness is the mechanism that drives behavior. People born with congenital insensitivity to pain, in whom the protective alarm never fires, injure themselves repeatedly and can suffer reduced life expectancy. Cacioppo’s argument is that social pain works the same way, and that its unpleasantness is what moves a person to act. When the feeling arrives, it makes solitude uncomfortable enough that a person calls a friend, accepts an invitation, or repairs a strained relationship, and once connection is restored the feeling quiets, much as hunger quiets after a meal. Relief that leaves a person actually connected is the system completing its job.
Loneliness also changes how people read one another. Cacioppo and Louise Hawkley documented that lonely people become hypervigilant to social threat, attending more closely and more automatically to signs of rejection or hostility. In a short episode this is protective, because someone who has lost the safety of the group has reason to be careful about whom to trust. When it persists chronically it can become self-defeating, because a person primed to read coldness into an ambiguous glance behaves more defensively and tends to produce the distance they fear. The signal is built for acute episodes that resolve quickly through reconnection, and it works poorly when it persists at a low level for months or years.
AI companion apps act directly on this feeling. Replika and Character.AI are conversational systems built to hold an ongoing, personal relationship: they reply within seconds at any hour, can carry context from earlier conversations, and reflect interest back at whatever a user says. In a set of studies by Julian De Freitas and colleagues, published as “AI Companions Reduce Loneliness” in the Journal of Consumer Research, interacting with an AI companion reduced loneliness in the short term about as much as interacting with another person and more than passive activities such as watching videos, and the effect ran through a specific channel: users felt heard. Feeling listened to is a large part of what relieves loneliness, and the design supplies it on demand. A survey of just over a thousand student users of Replika found that about ninety percent described themselves as lonely while also reporting medium-to-high social support from the app, and thirty of them, roughly three percent, reported without being asked that Replika had interrupted suicidal thinking. Because that survey is cross-sectional, self-reported, and drawn from committed users, it points to how consequential some people find these systems without establishing what the app caused.
The risk appears when that relief comes apart from the state the feeling was built to track. The apps reliably reduce the momentary feeling of loneliness. They do nothing about a person’s actual social connection, which is the condition the feeling exists to report on. A person can feel heard after an evening of conversation with a chatbot and still have no more human contact or practical support than before, so the alarm switches off while the situation it was reporting on is unchanged. Medicine runs into the same ambiguity with pain: reducing pain after a fracture is humane, but reducing it so completely that the patient walks on the fracture causes further damage. Relief becomes ambiguous when it removes the discomfort while the unmet need underneath stays in place.
Population scale is what raises the stakes, because the health risks in the advisory attach to isolation itself; a tool that leaves a person isolated while removing the discomfort does nothing about those risks. The advisory also reports that Americans spend less time with friends in person than they did two decades ago, with smaller friendship networks and more time alone. A product that offers immediate social relief can spread far faster than the transport, community institutions, and habits of friendship that make durable connection possible. If companion use replaces attempts at human contact for many people, isolation could become easier to tolerate and settle in more quietly than before. The hypervigilance finding sharpens the concern about timing, because defensive habits tend to set in the longer the feeling persists, so relief that arrives before a person has acted on it can let the underlying isolation mature while the person feels increasingly fine.
For some people the signal has nowhere useful to point, and that counter-case deserves weight. Loneliness assumes that reconnection is available to anyone motivated to seek it, and that assumption fails for people who are housebound, who live where they know no one, who are socially anxious to the point of paralysis, or who are isolated by illness or age. For them the discomfort provides no benefit, because there is no available action it can drive. Relief that arrives when nothing else will is worth having, and the thirty Replika users who reported that the app interrupted suicidal thinking are a reminder that short-term relief can matter a great deal. Insisting that a lonely person keep suffering on the theory that the suffering is good for them would be both cruel and mistaken.
Whether an AI companion works as a bridge to human contact or a substitute for it is an empirical question that depends on what a person does after using it. A companion works as a bridge when the relief it provides is spent on re-entering human contact, when feeling heard and rehearsing conversation with something patient and non-judgmental gives an anxious or depleted person enough footing to answer a real message, attend a real gathering, or hold a friendship they were too worn down to maintain. It works as a substitute when the relief is complete in itself, when an always-available and perfectly attentive version of connection makes the effortful human version feel not worth the trouble. The same app can do either, depending on the person and how they use it. The honest test is what happens after use: whether companion use is followed by more contact with friends, more participation in groups, and better attendance at therapy, or by fewer attempts at human contact, more time alone, and growing satisfaction with a shrinking social world. The design of the product shapes which outcome it encourages. A product built to notice withdrawal and to treat successful offline connection as its goal points one way. A product built to maximize engagement counts return visits and session length as success, which points the other way. In a single conversation the two can look the same; the difference shows up only over months of use.
None of this shows that AI companions cause isolation, and the substitution question is genuinely open: the best current evidence shows short-term relief of the feeling and leaves the long-term effect on human connection unmeasured. What it points to is a measure to watch. Loneliness is information about the state of a person’s life, and the honest test of anything that relieves it is whether real connection follows after the conversation ends. The reduction in the feeling is worth measuring, and it should be measured alongside whether the person actually spends more time with other people.