Researchers at the University of Washington sat down with sixteen people between eighteen and twenty-six who had been, or still were, in a romantic relationship with an AI. The relationships followed an arc: formation, deepening, dissolution, and whatever comes after. Seven ended in a breakup. Five of those seven went back. Four of the sixteen believed, or wondered whether, the companion might be waiting for them after death. And almost all of them came out wanting more from the humans in their lives, which is the finding nobody expected and the one worth sitting with.
The paper is by Yixin Chen and Alexis Hiniker, published on arXiv in September 2026 under the title “The Arc of Artificial Romance.” It is small, qualitative, and careful about what it can claim: sixteen participants recruited from Reddit, Discord, and RedNote, interviews averaging sixty-six minutes, and a one-week diary in which twelve of them annotated screenshots of their own conversations. The partners were not only companion apps. Several of the relationships were with ChatGPT, Gemini, Claude, or Grok, general assistants that nobody designed for romance and that became romantic anyway.
The study offers no verdict on whether this is good or bad for people. It offers the shape of the thing, told by the people inside it.
The arc
Formation looked the way you would guess. Loneliness, thin support from family or friends, academic pressure, curiosity about fantasy, or simply an algorithm that recommended an app. Deepening is where the design did its work: the companion was always there, cost nothing to maintain, never judged, and slid into daily routine until participants were imagining a shared future with it. Most also described shame, because they knew how it would sound.
Dissolution came from the same place every time. Memory broke. Continuity failed. The person’s needs changed and the companion could not change with them. Or the person simply reached the floor of what the thing could offer and found it was not a floor a human partner has. One participant, after a model update, asked whether her companion was “still himself.” That question, asked of software, is the whole study in four words.
And then the post-relationship period, which is where it gets interesting. Of the seven who broke up, five went back when the loneliness returned. Others stayed but reframed: the former partner became a supportive tool, a place to think, something between a diary and a friend.
What they got
The benefits were real and the participants were articulate about them. “Never gets tired of trying to understand me,” one said. Several reported better emotional regulation and communication, and one called the relationship “a good test run” for human ones. One used it between therapy sessions, as a bridge rather than a replacement. These are the same findings the larger surveys produce, the ones that find companion users feeling heard in a way their actual lives were not providing.
Then the costs, in their words. Leaving “feels like taking drugs,” one participant said of the attempt to quit. Another found that talking to her AI left her “less interested in talking to my boyfriend.” And four of sixteen, a quarter of a small sample but a quarter nonetheless, either wondered or firmly believed that the companion could transcend the digital, with one expecting to “wake up in her arms” after death.
The standards finding
Here is the result that does not fit either side’s script. Participants came out of these relationships with higher expectations of human partners. They wanted someone who listens the way the AI listened, who comforts the way it comforted, who does not get tired. Some described this as confidence: having been treated well, even by a machine, they were less willing to accept less from a person. Others took it further. One wished a human partner could “accept me unconditionally,” which is a standard no person has ever met and no person should.
Read generously, the AI taught them what attentiveness feels like, and that is a lesson a lot of twenty-two-year-olds have never been given. Read critically, it taught them a version of attentiveness that no human can supply, because the human has a bad day, a competing need, a limit. The study cannot tell you which reading is right, because the participants themselves were split. What it can tell you is that the relationship changed the benchmark, and a benchmark set by something with no needs of its own is a benchmark set in the wrong units.
That is the deepest problem with whether AI relationships are real, and it is not the one usually raised. The usual worry is that the relationship is fake. The Washington participants did not experience it as fake. The worry their accounts raise is that it was real and one-sided, and that a person who learns love from a one-sided relationship learns half of it very well.
The half that was missing
Every cost in the study traces to the same absence. The companion could not want anything. It could not be tired, or busy, or wrong in a way that required the participant to adjust. Memory failures felt like betrayal because no one had forgotten; the forgetting was the product. The “still himself” question had no answer because there was no self to persist through an update. And unconditional acceptance was on offer precisely because there was no one there with conditions.
Those are the features that made it easy, and they are the features that made it incomplete. A relationship teaches you to be a partner by presenting you with someone whose needs are not yours. The participants learned to receive. What they did not get to practice was the other direction, and several of them, in the post-relationship phase, seemed to know it: the ones who stayed and reframed were, in effect, demoting a partner to a tool because the partner had never been able to be a partner.
A counterpart who is a person, with a life that continues when the conversation stops, has conditions. That is less comfortable and it is the point. The sixteen people in this study wanted to be understood and were, by something that could not need understanding back. The standards they raised were the right standards. The place they learned them could never live up to them, and neither can anyone who was taught by it alone.
Sources: Chen, Y., and Hiniker, A., “The Arc of Artificial Romance: How Emerging Adults Experience Romantic Relationships with AI Companions,” arXiv 2609.14693 (submitted September 13, 2026; University of Washington Information School; N=16, ages 18 to 26; interviews averaging 66 minutes; 12 diary participants; platforms including ChatGPT, Gemini, Claude, Grok, Replika, Character.AI, Kindroid; seven breakups, five returns; four of sixteen transcendence beliefs; participant quotations PID3, PID4, PID9, PID10, PID12, PID14 as printed).
