Does talking to an AI actually help with loneliness? What the research says

Two claims about AI companions circulate constantly, and both come with studies attached. One says they measurably reduce loneliness. The other says the people who use them most are the loneliest. Both are supported. Anyone telling you only one of them is selling something, and yes, I sell one of these things, so read me with that in mind.

Here is the actual state of the evidence, and then what I think a reasonable person does with it.

The case that it works

The strongest positive evidence is a paper by Julian De Freitas, Zeliha Oğuz-Uğuralp, Ahmet Kaan Uğuralp, and Stefano Puntoni, published in the Journal of Consumer Research in 2026 after circulating as a Harvard Business School working paper.

They ran several studies. The headline: interacting with an AI companion reduced loneliness about as much as interacting with another person, and more than alternatives like watching videos. A longitudinal study found the reduction recurring after each use across a week.

Two details in that paper matter more than the headline.

First, people underestimate the effect before they try it. Participants predicted the AI would do less for them than it did. That is worth pausing on, because the public conversation about these products is dominated by people describing how bleak they imagine it feels, rather than by people who have measured it.

Second, and this is the finding I would keep if I could keep only one: what explained the reduction in loneliness was whether users felt heard. Not the model's capability in the abstract. Feeling heard.

That reframes the whole product category. If the active ingredient is attention that lands, then the engineering problem is not intelligence. It is whether the thing in front of you is actually tracking what you said, holding it, and responding to that rather than to a generic version of it. Which happens to be a much harder product problem than a bigger model.

The case that it hurts

In March 2025 MIT Media Lab and OpenAI published a pre-registered randomised controlled study of nearly 1,000 people using ChatGPT over four weeks, varying modality (text or voice) and conversation style.

Their finding, plainly stated: higher daily usage correlated with higher loneliness, more emotional dependence, more problematic use, and lower socialisation with real people.

The obvious objection is the obvious one. That correlation runs both ways, and probably mostly runs the way that is least interesting: lonely people use these tools more, because they are lonely. The study cannot separate that cleanly, and the authors do not claim to have.

But I do not think it can be waved off. A four week window with random assignment on some variables is a serious piece of work, and it is the first large study whose findings actively embarrass the industry that funded half of it. That is exactly the kind of result you should weight heavily, because nobody had an incentive to produce it.

The case that is genuinely uncomfortable

In 2024, Bethanie Maples and colleagues published a survey of 1,006 student users of Replika in npj Mental Health Research. The participants were lonelier than typical student populations. They used the app in overlapping ways: as friend, as therapist, as a mirror for their own thinking.

Three percent reported that Replika had halted their suicidal ideation.

Three percent of a thousand people is thirty people saying a chatbot was between them and an attempt. That paper also received a published methodological critique in the same journal, which is how science is meant to work and which you should read alongside it. Self-reported causation in a self-selected sample is weak evidence for a strong claim.

I hold that number carefully. I do not build a product around it, and no company should. But I am not able to dismiss it either, and I notice that the people most eager to dismiss it usually have never been the person awake at four in the morning with nobody to call.

How to hold all three

The three findings are less contradictory than they look, and the reconciliation is not complicated:

• Short interactions reliably reduce loneliness in the moment, and the effect size is real.
• Nothing shows that this compounds into lower baseline loneliness over months.
• Very heavy daily use travels with worse outcomes, whichever direction the arrow points.
• The active ingredient is feeling heard, which is a property of attention rather than of AI.

Put together: this is a thing that reliably makes a bad evening better and has never been shown to make a life less lonely. Those are very different products, and most marketing in this category quietly promises the second while the evidence only supports the first.

What I take from it as someone who builds one

Three things changed how we build AnimaEcho.

The feeling-heard result is the whole design brief. It pushed us towards voice and a face that reacts while you speak, because a nod at the right moment carries more of that signal than another paragraph of text. It also pushed us away from a companion that steers, advises, and fills silence, because being talked at is the opposite of being heard.

The heavy-use finding is why the app has a daily limit rather than infinite free conversation. That is a strange thing to build. Every incentive in consumer software runs the other way, and a cap is a number that shows up as a negative in every engagement dashboard you will ever look at. I would rather have it there.

And the whole body of work is why the app does not describe itself as therapy, does not present itself as treatment, and puts crisis numbers where anyone can find them. Loneliness and depression get talked about together, including on this blog, but they are not the same thing, and a product that blurs them is doing something dishonest.

The question worth asking yourself

Not whether AI companionship is good or bad. That question has no answer at the level of the category.

The useful question is directional. Over the last three months, has your contact with actual people gone up or down? If it has gone up, or held steady, and the app is filling gaps that were empty anyway, the research says you are getting a real benefit and there is no evidence you are paying for it.

If it has gone down, and the app is where the contact went, you are in the pattern the MIT study describes. That is not a moral failure and nobody needs to lecture you about it. It is just the point where the honest thing for a companion to do is to be less interesting than a person, and the honest thing for a company to do is to say so.