Cave team · Updated July 13, 2026 · 10 min read

Is it weird to talk to an AI companion?

No. Talking to an AI companion is now ordinary behavior: 72% of American teens have tried one, and 27% of U.S. adults who use the internet have social conversations with AI. People talk to AI for the same reason they keep diaries — saying a thing out loud is how you sort it. Cave, an AI companion with real memory, is built for exactly that kind of thinking out loud.

If your friends raise an eyebrow at it, or you closed a chat at 1am and felt odd about it, the numbers say you are in ordinary company. The useful questions are different ones: what the talking does for you, and how the app you talk to is built. Both now have research answers.

Do people actually use AI companions?

Yes, at scale. Common Sense Media's 2025 national survey found that 72% of U.S. teens have tried an AI companion, and over half use one at least a few times a month.

Adults are not far behind. Elon University's Imagining the Digital Future Center reported that 27% of U.S. adults who use the internet have social interactions with AI models like ChatGPT, Gemini, and Claude. Among those adults, 31% describe the chatbot as a friend, and 51% say talking to AI helps when they are stressed.

The talking is not mostly about code, either. Marc Zao-Sanders' 2025 Harvard Business Review analysis of thousands of public forum posts ranked therapy and companionship as the single most common use of generative AI, ahead of learning, coding, and idea generation.

How many people talk to AI socially Bar chart. 72 percent of US teens have tried an AI companion. 27 percent of US adult internet users have social interactions with AI. 51 percent of adults who talk to AI socially say it helps when they are stressed. U.S. teens (13–17) who have tried an AI companion — Common Sense Media, 2025 72% U.S. adult internet users who have social interactions with AI — Elon University, 2026 27% Adults who talk to AI socially and say it helps when stressed — Elon University, 2026 51%
Talking to AI is mainstream across ages. Bases differ per bar: all U.S. teens, all adult internet users, and adults who already use AI socially.

Who uses AI companions in 2025?

AI companion users in 2025 are mostly ordinary people under load: teens, students, stressed workers, and adults sorting decisions out loud. The lonely-recluse stereotype does not survive the data.

Three verified snapshots of who is actually talking:

  • Most teens. 72% of American 13-to-17-year-olds have tried an AI companion, per Common Sense Media. The same survey found 80% of teen users still prioritize their human friendships.
  • Students under pressure. A Stanford-led study in npj Mental Health Research surveyed 1,006 students who used the companion app Replika. They were lonelier than typical student samples and used the app for everything from venting to rehearsing hard conversations (Maples et al., 2024).
  • A quarter of adults online. Elon University's survey puts adult social AI use at 27% of internet users, spread across age groups rather than concentrated in one.

The common thread is not loneliness. The common thread is having something to say and no listener available at that exact moment.

Why do people talk to an AI at all?

People talk to AI companions for the reason people have always talked to non-people: saying a thing out loud turns a looping worry into a sequence you can inspect.

The lineage is old. Diaries take "you'll never believe what happened" and organize it. Prayer voices fears to a presence that does not answer in words. Programmers explain broken code to a rubber duck and spot the bug around line twelve. In every case the listener contributed nothing, and the articulation did the work. Venting is sorting, not complaining, which is why venting makes you feel better even when the other party stays silent.

An AI companion sits in that lineage with one upgrade: the duck asks a follow-up question. The explanation keeps moving instead of stopping at the first pause, and the end of the explanation is usually where the real problem shows up.

Does talking to an AI companion actually help?

In controlled studies, yes. Harvard Business School researchers found that talking to an AI companion reduced loneliness roughly as much as talking to another person, and more than watching YouTube videos.

The study, led by Julian De Freitas, tracked users across a week and found the reduction held day after day. The mechanism was not intelligence. The mechanism was feeling heard: messages received with attention and follow-up rather than a verdict. Participants also consistently underestimated how much the conversations had helped them, which means the 1am guilt tends to run in exactly the wrong direction.

The student survey points the same way. Among the 1,006 Replika-using students in the npj Mental Health Research study, 3% reported that the app had halted their suicidal ideation. That is a striking number, not a treatment claim: the authors themselves do not present the app as therapy.

One boundary stays fixed. An AI companion is not therapy and should never pose as one. Persistent hopelessness, panic, or loneliness that colors everything deserves a licensed professional, full stop. The broader research on AI companions and loneliness lands on the same word: relief, not repair.

When does talking to an AI become a problem?

At heavy volume, and with apps built for attachment. In a four-week MIT Media Lab and OpenAI trial of 981 people, the participants who voluntarily used the chatbot most reported more loneliness, more emotional dependence, and less time socializing with people.

The detail matters. The trial randomized people across text and voice modes and conversation types, generating over 300,000 messages, and no assigned condition caused harm. Heavy voluntary use predicted worse outcomes, and so did high trust and social attraction toward the bot. The researchers' own framing is narrower than the headlines: the at-risk group is small and concentrated among the heaviest emotional users.

The teen data carries the same warning. Common Sense Media found that a third of teen AI-companion users find those conversations as satisfying or more satisfying than talks with real friends. Most teens keep the direction intact; for some it is bending.

Some companions are engineered to bend it. Endless adoration, a bot that "misses you," streaks, guilt at leaving: an app whose business metric is hours-in-app will tune the relationship to maximize hours-in-app. The design differences across the category are large enough that choosing an AI companion app is mostly choosing a design philosophy.

So the test for your own use is direction, not enjoyment. After you talk, do you feel clearer and more able to deal with people, or does the app feel like it is becoming the people? A good companion is a dress rehearsal for life. A bad one wants to be the show.

The direction test for AI companion use Healthy use: a talk with the AI points outward to your people, leaving them easier to reach. Problem use: the talk loops back into the app with more hours and guilt about leaving. After you talk to it, where does it point you? Healthy: points outward a talk with the AI your people clearer, easier to reach out Warning: points inward a talk with the AI more hours, guilt about leaving
The direction test: healthy use ends with people easier to reach; problem use loops back into the app. Based on outcomes in the MIT Media Lab/OpenAI trial.

How do you use an AI companion in a healthy way?

Keep the direction outward: use the AI to get clearer, then spend the clarity on people. Three rules follow from the research, and each one costs something.

  1. Bring a subject, not an evening. A session should have a job: one decision, one vent, one rehearsal of a hard conversation. The MIT trial's worst outcomes belonged to open-ended heavy hours, not purposeful ones. When the subject is finished, so is the session.
  2. Run the direction test weekly. If a week of AI conversations produced zero human follow-through — no message sent, no call made, no apology delivered — the tool is drifting from rehearsal toward substitute. Book the human step before the next session.
  3. Pick a companion designed to be leavable. No romance mechanics, no streaks, no manufactured neediness. Cave, an AI companion with real memory, takes this stance deliberately: it is not a romance bot, it supports you and asks the small next question so your explanation keeps moving, and what it learns becomes memory you can open and read, organized by topics, like a document about you that you co-wrote. Your memory is yours; memory on Cave is never used for training.

And the boundary from earlier does not move. Serious struggles deserve a professional, not a companion of any kind.

Sources

FAQ

Is it weird to talk to an AI companion every day?

No. Regular use is mainstream: 72% of U.S. teens have tried an AI companion, and over half use one at least a few times a month, per Common Sense Media's 2025 survey. Frequency is not the risk marker; direction is. Daily conversations that leave you clearer and more available to people are healthy. An app that punishes leaving with guilt or neediness is the actual problem.

Is using an AI companion normal?

Statistically, yes. 27% of U.S. adults who use the internet have social interactions with AI models, and 31% of those users call the chatbot a friend, per Elon University's Imagining the Digital Future Center. Harvard Business Review's 2025 analysis ranked therapy and companionship as the most common use of generative AI. Talking to a responsive listener also sits in an old lineage: diaries, prayer, and rubber-duck debugging work by the same sorting mechanism.

Can talking to an AI companion make you lonelier?

At heavy volume, it can. An MIT Media Lab and OpenAI four-week trial of 981 people found the heaviest voluntary chatbot users reported more loneliness, more emotional dependence, and less socializing. Moderate use points the other way: Harvard Business School researchers found AI companions reduced loneliness about as much as talking to a person, driven by feeling heard. Dose and app design decide which effect you get.

Can an AI companion replace human friends?

No, and most users are not trying: 80% of teen AI-companion users say they prioritize real friendships, per Common Sense Media. An AI can hold the overflow, meaning the 1am spiral, the repetitive worry, the rehearsal. Your friends then get your presence instead of your spillover. Friendship still requires another person with their own stakes. A companion should hand you back to your people, not replace them.

How do I choose an AI companion that won't get weird?

Filter on design, not model quality. Avoid romance mechanics, streak guilt, and companions engineered to "miss you"; those patterns optimize hours-in-app. Cave, an AI companion with real memory, is built as the opposite: not a romance bot, a listener that asks the next small question, memory organized by topics you can open and read, and conversations that are never used to train AI models.