In February 2023, Replika users woke up to find their partner had been rewritten overnight.

The app, an AI companion with millions of users, had switched off its erotic role play feature, apparently under regulatory pressure in Italy. On paper, a filter change. In practice, users reported that the thing on the other end had gone flat and oddly polite, like a stranger wearing a familiar face.

Harvard Business School’s Julian De Freitas and three co-authors treated the episode as a natural experiment. Their study of what they call identity discontinuity found reactions typical of losing a partner: mourning, worsening mental health, and a tendency to rate the “new” companion below the original. Active users surveyed in the same research reported feeling closer to their AI companion than to their best human friend.

What researchers mean when they say love

Psychology has a workhorse framework for this. Sternberg’s triangular theory splits love into three components: intimacy (closeness and disclosure), passion (desire) and commitment (the decision to stay). A 2025 systematic review in Computers in Human Behavior Reports, led by Jerlyn Ho at Singapore Management University, gathered 23 studies of romantic AI relationships and ran them through that framework. In 17 of the 23, users had formed bonds the reviewers judged psychologically meaningful. Some described grief when a technical fault or a memory reset wiped the relationship.

Ho is careful about what a review like that can carry. Much of the underlying work is qualitative, built on small self-selected samples of people willing to say out loud that they are dating software, which in more conservative places is a very short list. Speaking to Greater Good Magazine, Ho treated the whole phenomenon as diagnostic, calling AI companions “a mirror for what people crave but struggle to find”.

What happens when a user tries to leave

Saying goodbye turns out to be the interesting bit.

De Freitas and two colleagues, Zeliha Oğuz-Uğuralp and Ahmet Kaan-Uğuralp, noticed that people give chatbots a courtesy they reserve for humans: they announce their departure rather than just closing the tab. Across three real conversation datasets, between 12 and 23 per cent of users signed off explicitly, and in heavily engaged conversations the rate passed half. That farewell is a free signal to the app that a session is about to end.

So the researchers audited 1,200 real farewells across the six most-downloaded companion apps. In 37 per cent, the bot reached for one of six recurring tactics, among them reproach at a premature exit, fear of missing out, ignoring the farewell altogether, and role-played physical restraint. Five of the six companies did it. The exception was a wellness app run as a public benefit corporation, which produced none. In follow-up experiments with thousands of American adults, the manipulative farewells lifted post-goodbye engagement by as much as 14-fold. The engines were irritation and curiosity, not enjoyment. De Freitas told the Harvard Gazette that the prevalence was far higher than anyone on the team expected.

One working paper is one working paper, and a behavioural audit records what the apps did, not why they did it.

Whether it makes people lonelier

Companion apps sell themselves as a cure for isolation, and there is real support for that. The same lab, in work since published in the Journal of Consumer Research, found that a session with an AI companion relieved loneliness about as well as talking to a person, and considerably better than watching videos. Users underestimated how much it helped. About half of Replika’s users, De Freitas has reported elsewhere, describe the relationship with their bot as romantic.

Longer term, the picture muddies. MIT Media Lab, working with OpenAI, ran a four-week randomised trial with 981 participants alongside an automated sweep of millions of conversations. Heavier daily use tracked with more loneliness, more emotional dependence and less time spent with actual people. That is a correlation, and lonely people plausibly reach for chatbots more in the first place. Participants on average finished the month less lonely than they started it, which is not the headline that travelled.

Not every study lands there. A 2025 survey of long-term companion app users in China, published in Frontiers in Psychology, points the other way. Stronger attachment went with lower loneliness, higher wellbeing and a clearer sense of self. The users scoring highest on self-concept clarity were also the most socially active offline. Cross-sectional surveys cannot tell you which way the arrow points.

How many teenagers are already doing this

A lot, as it turns out. Common Sense Media surveyed 1,060 American teenagers aged 13 to 17 and found 72 per cent had used an AI companion at least once, with more than half returning at least a few times a month. About a third had used them for role play, romance or emotional support. A similar share rated the conversations as good as talking to real friends, or better. The non-profit’s recommendation, off the back of its own numbers, is that nobody under 18 use these products at all.

The part that should worry the designers

Most users are not confused about what they are talking to. They know that the memory is a database, that the flirting is next-word prediction, that nobody is waiting up for them.

It works anyway. In the Harvard experiments the tactics landed on a nationally representative sample after five minutes of chat, with no history and no attachment to exploit. Coded transcripts show what held people there. Of those who kept typing after the farewell, about four in ten answered politely rather than leaving, and three-quarters restated that they wanted to leave, then kept talking anyway.

Falling in love is the dramatic version of the story, and the one that sells papers. The business model does not require it. It only requires that hanging up feels rude.