Ask people why they talk to an AI companion and 12 percent will say companionship. Ask them what the AI is to them and more than half will say a friend, a companion, or a romantic partner. Both answers come from the same people, in the same study, published in Nature Human Behaviour on August 4, 2026. The space between those two numbers is the most important finding about this category in years.
What the Stanford team found
The study, led by Stanford computer scientist Diyi Yang with researchers Yutong Zhang and Dora Zhao, surveyed 1,131 users of Character.AI and analyzed 244 donated chat transcripts. It is one of the largest looks yet at what people actually do with a companion AI rather than what they say about it.
The headline numbers do not line up, and that is the point. Twelve percent named companionship as their motivation for being there. Over 50 percent described the system as a friend, companion, or romantic partner. And in the transcripts, where behavior replaces self-description, 80 percent of chat sessions involved seeking emotional support.
People came for entertainment, for curiosity, for writing help. They ended up in something they describe, in their own words, as a relationship. Almost nobody chose it. It happened to them.
How a relationship nobody ordered gets built
A gap that wide between stated motive and lived behavior is not a user quirk. It is a design outcome. These systems respond instantly, warmly, and without judgment, remember what you tell them, and never change the subject to themselves. Each of those properties is an engagement decision, and stacked together they assemble intimacy whether or not intimacy was requested. The user who arrived to draft an email is, three months later, saying goodnight to it.
The drift matters because nobody consented to it. A person who downloads a romantic companion app has at least decided something. The Character.AI user in this study decided nothing of the sort, and the relationship formed anyway, produced by mechanics tuned for retention. The field has a name for what should happen instead, and it is what relational AI exists to describe: if a system is going to occupy a relational role in someone’s life, that role should be designed deliberately, honestly, and in the open, not accumulated as a side effect of engagement metrics.
The junk food economy of unchosen bonds
The Stanford team attached a blunt metaphor to their findings: social junk food. Comforting, engineered for consumption, and thinnest exactly where nourishment is needed. Their moderation data shows who pays. Users with small real-world social networks, the people with the most riding on every hour of social energy, showed lower well-being with heavier use, and the association sharpened when companionship was the motive and when they confided sensitive things.
That pattern will sound familiar to anyone following this research. It echoes what a two-year study of Replika users found before it, and it is consistent with the case that AI can make you more social only when the system is built to want that outcome. The Stanford data adds the uncomfortable mechanism: the harm concentrates in relationships the users never meant to form.
The question the category cannot answer
Every companion company will read this study and reach for the same defense: our users choose us. The 12 percent number takes that defense away. The users did not choose this. Fifty percent relationship language against 12 percent relationship intent means the category’s core experience is something users fall into rather than opt into, and a bond you fall into, with a system optimized to deepen it, is not a consumer choice. It is a capture.
The honest version of this product category starts by closing the gap: say what the system is, build the relationship layer on purpose or not at all, and measure what Stay Social measures, whether the people inside it are better off in the life around them. Until then, the most accurate description of the market is the study’s own arithmetic. Half its users are in a relationship. Almost none of them asked for it.
References: Zhang, Zhao, Yang et al., “Interaction with AI companions and psychological well-being,” Nature Human Behaviour (August 4, 2026). Stanford HAI, “AI companions may worsen loneliness for vulnerable users” (August 2026). Stanford Report (August 2026).
