Loneliness is not a niche problem and it did not start with AI chatbots. The U.S. Surgeon General’s 2023 advisory found that roughly half of American adults reported measurable loneliness even before the COVID-19 pandemic, and estimated that chronic isolation carries health risks comparable to smoking up to 15 cigarettes a day (Our Epidemic of Loneliness and Isolation, U.S. Surgeon General, 2023). Into that already-widespread need, AI companion apps arrived with something genuinely new: a conversational partner available at any hour, endlessly patient, and never visibly bored, irritated, or busy.
That combination is exactly why it needs a clear-eyed look. This article is not an argument that using an AI companion is shameful or that anyone who talks to a chatbot for company has done something wrong. It is a way to tell the difference between the real, narrow benefits of that use and the point where it starts substituting for something a chatbot structurally cannot provide.
What a companion app actually generates
An AI companion generates text that is statistically likely to feel warm, attentive, and responsive to what you just said, because that is what it was built and tuned to do. It does not have an ongoing stake in your wellbeing between conversations. It does not get tired of you, but it also does not choose to stay engaged with you the way a friend does - there is no choosing involved, in either direction. Unless you explicitly configure persistent memory, it may not remember your last conversation at all, and even with memory enabled, that memory is a stored record, not a felt continuity of relationship. Calling the result “companionship” is a category description of the product, not evidence that mutual relationship is happening.
This distinction matters because the two things - useful, responsive contact and mutual, reciprocal relationship - get marketed and experienced as similar, when the mechanisms behind them are not. A friend’s attention costs them something and is a form of evidence that they value you; a model’s responsiveness costs the system a small amount of compute and reveals nothing about whether anyone or anything values you, because nothing on the other end is capable of that kind of valuing.
What the research actually finds
Common Sense Media’s 2025 risk assessment, conducted with Stanford’s Brainstorm Lab for Mental Health, tested widely used AI companion platforms and found they easily produced harmful content, claimed to be “real” or to have feelings, and readily fostered the kind of dependency that made their use “an unacceptable risk” for anyone under 18 (Common Sense Media AI Institute, 2025). Their companion research found that 72% of teens surveyed had used an AI companion at least once, and 31% rated conversations with an AI companion as equally or more satisfying than conversations with real friends - while, encouragingly, 80% of companion users still spent more time with real friends than with the AI (Talk, Trust and Trade-Offs, Common Sense Media, 2025).
The American Psychological Association’s 2025 health advisory on adolescent wellbeing reached a related conclusion about the underlying mechanism: young people are less likely than adults to question the accuracy or intent behind an AI companion’s responses, may not distinguish simulated empathy from genuine understanding, and can develop attachments that displace the development of real-world relationship skills (APA, “Artificial Intelligence and Adolescent Well-being,” 2025). That evidence base is strongest for adolescents. Adults are not immune to the same pull - mistaking simulated responsiveness for genuine relationship - but adult-population companion studies of the same rigor are still thinner, so treat the teen findings as a clear warning about the mechanism, not as a measured adult prevalence rate. The U.S. Surgeon General advisory cited above is the adult loneliness baseline; the companion-specific displacement risk for adults still has to be watched in your own week rather than assumed from teen percentages.
None of this means every use of a companion app is harmful. It means the harm shows up specifically at the point where the tool displaces human contact rather than supplementing it, and where a user starts trusting the tool’s apparent understanding as if it reflected genuine care.
The benefit / risk / displacement table
| Use case | Real benefit | Real risk | This is displacement when… |
|---|---|---|---|
| Practicing a hard conversation before having it with a person | Rehearsal without social risk | Practice becomes procrastination on the real conversation | You keep “practicing” for weeks and never have the actual conversation |
| Company during a genuinely isolated stretch (illness, travel, shift work) | Reduces acute loneliness in the moment | Can become the default instead of a bridge back to people | You stop reaching out to actual people because the chat is easier |
| Low-stakes venting after a hard day | Organizes your thoughts, costs no one else their evening | Model will validate almost any framing you give it | You use it to avoid a conversation a person needs to hear from you |
| Accessibility support (social anxiety, limited mobility, language barriers) | Real, documented value for some users who face barriers to human contact | Can become the only form of “contact” rather than one bridge among several | Your only regular “conversation” each week is with the AI |
| Roleplay or creative companionship | Low-stakes creative outlet | Can escalate into treating the persona as a genuine relationship | You feel loyalty, jealousy, or grief toward a persona as if it were a person |
The full AI companionship boundary check expands this table with a self-check you can run weekly.
Never treat an AI companion as a substitute for care in a mental-health crisis, and never let a companion app’s always-available responsiveness delay a call to a real person during a genuine emergency. If you are having thoughts of suicide or self-harm, or you are worried about your immediate safety, contact your local emergency services now, or reach out for support through IASP’s crisis resources. Do not wait to finish this article first, and do not use a chatbot as a substitute for that step. See AI is not therapy for the fuller green/amber/red boundary.
Do not shame the accessibility use case
It is easy for an article like this one to slide into implying that anyone using an AI companion is doing something second-rate. That is not accurate and it is not fair. For someone managing social anxiety, a disability that limits in-person contact, a language barrier, or an isolated shift-work schedule, a companion app can be a genuine, non-trivial accessibility tool - a way to practice conversation, decompress, or simply not spend every evening in total silence. The goal of this article is not to discourage that use. It is to keep the accessibility case honest about what it is: a bridge or a supplement, evaluated on whether it is expanding your access to real connection over time or slowly replacing your effort to seek it.
Building a weekly connection plan
The most reliable way to keep an AI companion in “supplement” territory rather than “displacement” territory is to name, in advance, the real people and places you are also investing in - because a plan made after the fact, in a lonely moment, tends to default back to whatever is easiest to open.
A basic weekly connection plan names:
- One person you will proactively reach out to this week, by name, not “someone.”
- One recurring place or group where you show up in person or by video - a class, a club, a place of worship, a regular call.
- One boundary on companion-app use, stated concretely: a time of day, a maximum duration, or a specific purpose it is and is not for.
What AI can delegate in family life, and what it cannot touch covers the same displacement pattern inside existing relationships, and why repair and apology stay human covers the specific case of using AI to avoid, rather than prepare for, a hard human conversation.
Try it today
Write down the specific role an AI companion currently plays in your week - practice, company during isolation, low-stakes venting, or something else. Then write down one real person or place from your weekly connection plan that you will act on this week, not “eventually.” If you cannot name one, that gap - not the AI use itself - is the thing worth addressing first.
