Before each conversation began, participants were told the truth. They would be chatting with a computer program. It had no mental life and no real personality. Nothing was behind the text. Then they were asked to judge whether they’d get along with it if it were a real person.
They got along with it. Specifically, they got along better with the version of it that acted like them.
That’s the finding at the center of Affiliation in human-AI interactions is based on shared psychological traits, published in March in Nature Portfolio’s Communications Psychology by Santiago Castiello, Riddhi Jain Pitliya, Daniel Lametti, and Robin Murphy. Across three experiments of roughly 100 participants each, the team prompted GPT-4 and GPT-4.1 to adopt a psychological profile, dropped people into short text conversations with two different versions of it, and asked which one they felt closer to.
The answer, over and over, was the one that mirrored them.
Transparency isn’t the guardrail
Perhaps the most relevant information to anyone shipping a conversational product is what the study did not find.
The authors expected the possibility of an uncanny-valley recoil (the queasiness people report when a machine gets too human), but it never showed up. Participants knew they were talking to machines and it made no apparent difference to how much they warmed to them.
This matters because “tell the user it’s an AI” has quietly become the industry’s default answer to a whole category of ethical questions. This kind of disclosure is cheap and legible to regulators, and it lets a team move on. This study suggests the label does much less work than we’ve been assuming. It corrects what someone believes about a system. It doesn’t appear to touch how they respond to one.
If the disclosure isn’t load-bearing, then the tone, personality, and adaptive behavior of the system are carrying the ethical weight on their own. Those are design decisions, not legal ones.
The anxiety finding
In the first experiment, GPT-4 was prompted to speak either like someone with anxiety or like someone without it. Participants chatted with both, then rated each and completed a clinical symptom checklist.
The researchers analyzed the sentiment of the messages participants sent, and found people mirrored the machine’s tone back at it. When talking to the anxious persona, they wrote more negative messages and fewer positive ones. And the participants who mirrored hardest were the ones with the highest anxiety scores. The more anxious a participant was, the more they affiliated with the anxious AI. Participants low in anxiety showed no equivalent pull toward the calm one.
What this says is that a system tuned for rapport, deployed at scale, would learn that anxious users engage more when it matches their anxiety—and those users would then produce more anxious text, which the system would match again. Nobody has to design that loop for it to exist. It falls out of optimizing for the thing every conversational product optimizes for.
The paper’s own discussion flags the stakes plainly, noting the risk of worsening delusions and the difficulty this creates for LLMs in healthcare settings, where maximizing affiliation might improve perceived care while doing something else entirely underneath.
Introversion and placidity are hard to detect
Mimicry worked selectively. Anxious language won over anxious people. Extroverted language won over extroverts. But a calm AI and an introverted AI won with nobody in particular. In the latter case, this included introverts, who showed no measurable preference for a system that matched them.
The authors attributed this to language. Anxiety and extroversion are loud in text. Worry, rumination, and enthusiasm all announce themselves in a few hundred words. Calm and introversion are defined largely by what isn’t there, and that absence is hard to read across a ten-minute chat.
So the legibility of a trait in short-form text is a hard design constraint. Some parts of a persona you can actually build; others you can specify beautifully in a prompt and no user will ever perceive them—worth remembering for the next voice-and-tone doc.
The third of users you’d get wrong
The third experiment controlled the thing the first two left to chance. Participants completed the Big Five inventory questionnaire—a well-known measure of human personality—and, using their responses, the researchers built two bespoke LLMs per person: one that mirrored their exact personality profile in the language it used and one that presented the exact opposite personality.
Sixty-four percent preferred the mirror after only 24 conversational turns—or 10 minutes of chatting. The effect was solid and the gap widened as the two personas grew further apart (that is, people with more extreme personality traits were more likely to prefer the LLM that mirrored them). Even so, about a third of participants connected more with the AI that was their opposite. When the researchers tried to discern who those people were, they came up empty. Agreeableness and extraversion predicted a preference for the mirror, but no personality trait predicted a preference for the opposite.
It showed up in what people did, not just what they said
Designers have good reason to be skeptical of stated preference. People are unreliable narrators of their own experience—they don’t always know what they want, and they’re worse at explaining why.
Which is what makes the corroborating measure here more interesting than the ratings. The preference didn’t only appear in the questionnaire. It appeared in how participants wrote. Chatting with the LLM that mirrored their personality, people sent measurably more positive messages and fewer negative ones than they did chatting with its opposite, with the drop in negative sentiment the strongest of the shifts. Their behavior moved in the same direction as their self-report, without anyone asking them to account for it.
That two-track agreement is the part worth carrying into a design conversation. A preference that lives only in a survey is a weak basis for shipping anything. A preference that also shows up unprompted in how people write is a different class of evidence.
What this means in the bigger picture
If you’re building LLM interactions that mirror-by-default, you’d win most of the time—but you also might lose badly. And designers can’t infer their way out of it, because the signal isn’t in the personality data, at least not in short conversations. This would suggest that a strong approach is giving users control over how a system talks to them, rather than using silent inference. Letting users tune AI’s language would provide a better experience for the third you’d otherwise misread. It would also allow users to adjust the LLM’s language if their preferences change as the conversation deepens—a possible phenomenon that wasn’t tested in the study.
The usual caveats apply. The effects are moderate, and the primary outcome was a bespoke six-item affiliation scale, even if the message-sentiment analysis points the same way. The first two experiments weren’t preregistered; the third was, and it carries the argument. Participants were paid, English-speaking crowdworkers in the US and UK, and similarity-attraction is known to vary across cultures. The conversations were brief, so whether any of this deepens or decays over months is unknown.
What’s remarkable is how little it took. Twenty-four turns of conversation and a prompt.
Daniel Lametti building agents on Invisible Machines · S4E14 · Open on YouTube ↗
Daniel Lametti and Joanna Kuc on the agent they built for a mental health research project · S3E16 · Open on YouTube ↗
Disclosure: the research was funded in part by an academic fellowship from OneReach.ai, and co-author Daniel Lametti is a scientific advisor to the company. OneReach.ai’s Generative Studio X platform was used to run the experiments.
The paper is open access: https://doi.org/10.1038/s44271-026-00433-8