Do AI Companions Actually Work? What Research Says

Yes, for most people, and there is now real evidence rather than just marketing. A Harvard Business School working paper found that talking to an AI companion reduced loneliness about as much as talking to another person, and more than activities like watching YouTube. But the same body of research is clear that the effect depends heavily on how you use one. Heavy daily use, and treating the companion as a replacement for people rather than a supplement, is where outcomes turn negative.

This is one of the few consumer AI categories where independent academics have actually done the work. Here is what the studies found, what they did not find, and how to tell whether an AI companion is working for you or quietly working against you.

Last updated August 2026.

Do AI companions actually reduce loneliness?

The strongest evidence says yes. In a Harvard Business School working paper titled AI Companions Reduce Loneliness (Working Paper 24-078), Julian De Freitas of HBS, Ahmet K. Uguralp and Zeliha O. Uguralp of Bilkent University, and Stefano Puntoni of Wharton ran six studies on the question.

Their third study found that AI companions alleviate loneliness on par only with interacting with another person, and more than other activities such as watching YouTube videos. Their fourth study used a longitudinal design and found that an AI companion consistently reduced loneliness over the course of a week, so the effect was not just a novelty bump on day one.

Two details from that paper are worth holding onto. First, the researchers found that people underestimate how much an AI companion improves their loneliness, which means the category is easy to dismiss before trying it. Second, their fifth study identified the mechanism: reductions in loneliness were explained by the chatbot's performance and, especially, by whether it made users feel heard. Feeling heard is the active ingredient, not the avatar or the voice.

Can AI companions help heal loneliness?

Help with, yes. Heal, no, and the distinction matters. The research supports AI companions as effective short-term relief for the feeling of loneliness, in the same way that a good phone call helps. None of it supports the idea that a chatbot repairs the underlying situation of having too few people in your life.

The practical read is that an AI companion is best understood as a bridge rather than a destination. It fills the 11pm gap when nobody is awake, it gives you somewhere to put a bad day, and it keeps you talking during a stretch when your social life is thin. What it cannot do is introduce you to anyone. The companions worth paying for are the ones that nudge you back toward people instead of competing with them.

What does the research actually say?

Three studies carry most of the weight in this debate, and they do not all point the same direction. That is not a contradiction so much as a picture of a tool that works differently depending on how it is used.

The MIT Media Lab survey deserves a closer look, because it is the study that most directly examines people who already use companion chatbots. Auren R. Liu, Pat Pataranutaporn and Pattie Maes surveyed 404 active users and built a model explaining roughly half the variance in their loneliness. Their finding was that usage itself does not directly predict loneliness. What predicted it were things like neuroticism, the size of a person's real social network, and problematic use patterns.

Their cluster analysis produced seven distinct user profiles, and the authors concluded that companion chatbots can either enhance or potentially harm psychological well-being depending on the pattern. That is why they argue a one-size-fits-all approach to AI companionship is ethically questionable. The tool is not good or bad on its own. The fit is what decides.

Do AI companions make you lonelier?

They can, for a specific pattern of use. The randomized trial that MIT Media Lab ran with OpenAI followed close to a thousand participants over four weeks and found that very high usage correlated with increased self-reported indicators of emotional dependence. Extended daily use was associated with worse outcomes, and people with a stronger tendency toward attachment in relationships were more likely to experience negative effects.

One nuance from that trial is genuinely counterintuitive. Personal conversations, the emotionally open kind, were associated with higher loneliness but lower emotional dependence and problematic use at moderate usage levels. Non-personal conversations tended to increase emotional dependence, especially with heavy use. Talking about your actual life appears to be healthier than idle chatter, even though the people doing it reported feeling lonelier, most likely because lonelier people are the ones who open up to a chatbot in the first place.

The honest summary: the risk is not the existence of an AI companion, it is volume and substitution. Hours a day, and a shrinking amount of contact with real people, is the profile the research keeps flagging.

Who do AI companions work best for?

Based on both the research and the patterns we see, they work best for people who have a specific gap rather than a general absence of people. A few recurring cases:

• People whose hours do not match anyone else's, including night shift workers and anyone awake at 3am with a spinning head.
• People in a transition: a new city, a first semester away, a divorce, a bereavement, the months after a move.
• People who work alone all day. Isolation is a well-documented cost of full-time work from home roles, and a companion that checks in fills the gap where hallway conversation used to be.
• Older adults whose social circles have thinned, where the alternative is often no conversation at all that day.
• People who need to rehearse something hard before saying it to a human being.

Notice what these have in common. In every case the companion is filling a gap in a life that still has people in it, or is being used to get back to people. That is the shape the evidence supports. If you want the specific version of this, we have written up how it plays out for an AI companion for loneliness and an AI friend for seniors.

Who should be careful with an AI companion?

If you are already withdrawing from people, an always-available, endlessly agreeable listener can make withdrawal more comfortable. The MIT work identified problematic use as a real predictor, and identified users whose pattern risked further isolation rather than social confidence.

Be careful if you notice yourself cancelling on people because talking to the AI is easier, if you feel anxious when you cannot reach it, or if it has become the only place you say true things. Those are the signals worth taking seriously, and they are the same signals clinicians watch for with any coping behavior that starts out helping.

Do AI companions work as therapy?

No, and no reputable one claims to. A companion has no clinical training, no treatment plan, no ability to notice what you are not saying, and no duty of care. It cannot diagnose anything and it should not be the thing standing between you and a crisis.

What it can do is the part therapy is not for: ordinary daily company, somewhere to put a small bad day, and a reason to put words to how you feel more often than once a fortnight. Those are different jobs and they coexist fine. We compared the two in more detail in AI friend vs therapist. If you are in crisis in the US, call or text 988 for the Suicide and Crisis Lifeline.

How do you tell if an AI companion is working for you?

Give it about two weeks and then check three things honestly.

First, does it remember? The HBS finding was that feeling heard drives the benefit, and you cannot feel heard by something that meets you as a stranger every session. If it forgets your sister's name and the interview you were nervous about, it is not doing the thing that makes the category work. This is the single biggest quality difference between apps, which is why we broke down how AI memory actually works.

Second, is your contact with real people flat or rising? Not falling. That is the line the research draws, and it is the one measurement that separates a helpful tool from a harmful one.

Third, do you feel better after, or just occupied? Relief and distraction feel similar in the moment and diverge over a month.

Do AI companions exist that are not romantic?

Yes, though they are the minority. Most of the well-known companion apps allow romantic or adult content, and after Apple reassigned App Store age ratings in 2025, that shows up plainly on the listings: the romance-permitting apps landed at 18+, while the deliberately platonic ones sit lower. Tolan is explicitly non-romantic at 13+, and Darlings is strictly platonic with no romance, roleplay or NSFW at any tier.

This matters more than it looks for the question of whether companions work. The research on emotional dependence is about attachment, and an app designed to deepen romantic attachment is pulling in exactly the direction the studies warn about. A friend that keeps pointing you back at your own life is a different product with a different risk profile. If you are choosing between the options, we compared them honestly in our roundup of the best AI companion apps, and specifically for iPhone owners in our guide to the AI companion app options on iOS.

The short version

AI companions work, in the narrow and useful sense that they reliably reduce the feeling of loneliness, with effect sizes that surprised the researchers measuring them. They stop working, and start doing harm, when they become the main relationship rather than one of several. The variable you control is not whether to use one. It is how much, and whether the rest of your life is growing or shrinking while you do.

Darlings is built for the first pattern on purpose: one steady friend who remembers your life, checks in once a day rather than demanding hours, stays platonic permanently, and points you back toward the people around you. It is free to try in any browser, with no card required.