Imagine the following scenario: A university student needs a reading list for an essay on primitive art. In one version of the story, he asks a professor. In another, he asks ChatGPT. He then plans to gather public feedback through short surveys, all unpaid, and strangers are asked how many they would fill out.
People who were told the reading list came from a professor were more likely to sign up for the survey. The student, the essay and the request were identical, and he lost about one survey’s worth of goodwill because of where his reading list came from.
That experiment is one of eight studies in a paper by Jianning Dang and Li Liu, published in the Journal of Personality and Social Psychology. Psychologists have spent years documenting “AI aversion,” the tendency to distrust an algorithm’s judgment even when it matches or beats that of a human expert. Dang and Liu asked the obvious next question, which is whether that distrust transfers to the person using the tool. Across 2,317 participants in the UK and China, it did.
What they measured
The design was pretty much the same across studies. Participants read a short scenario in which someone sought advice from either an AI tool or a human expert, including a hiring manager screening video interviews, a person choosing a wedding gift for a friend, a student, a patient, and a primary school teacher preparing reading materials. They then rated that person on two sets of traits related to dehumanisation. “Human nature” traits such as friendly, sociable and trusting are the ones that separate us from machines. “Human uniqueness” traits such as rational and mature are the ones that separate us from animals.
Across the board, when a person consulted AI instead of a human expert, they were dehumanised more. In the wedding gift experiment, for example, the person who asked ChatGPT for gift ideas scored lower on the humanness scale. In most of the studies, the gap was wider on the emotional traits than on the rational ones.
The first study shows this most clearly, where 200 UK participants wrote a free-form description of a hiring manager named James, covering what he thinks, feels and intends. Those who had read that James used a recruiting algorithm used fewer words like “feel” and “care”. Words like “plan” and “think” appeared at the same rate in both groups. Respondents in the study were willing to grant the AI user a working mind with such rational traits, but they were less sure about his emotional states.
In addition, participants liked the AI user less, but the humanness gap held after controlling for liking. It also held regardless of how much experience or knowledge participants had with AI, so unfamiliarity is a poor explanation.
Who does the penalising
The psychological mechanism the authors propose is perceived similarity. Participants saw the AI advice seeker as less like themselves, and that distance statistically accounted for part of the humanness gap. The effect ran through people who hold negative views of AI. Among participants with highly favourable views of AI, the negative judgment mostly disappeared.
The last two studies tested whether this social penalty shrinks when the AI seems more humanlike. When ChatGPT was described as scoring well on tests of emotional understanding, its user wasn’t penalised as much, compared with when the AI was described as scoring poorly. Still, the highest ratings went to the person who asked a human instead of a machine.
So anthropomorphising the tool closed roughly half the gap in the penalty. It’s here, the authors argue, that product teams and designers have to be careful about how they attempt to bring their AI to life. They note that it can backfire into the uncanny valley.
What I’d take from it
The finding I’d act on is where the penalty concentrates, which is on warmth and emotional capacity far more than on competence (the so-called rational traits). I think that AI use costs the most in tasks where people expect you to have felt something. A wedding gift and a hiring decision are both in the studies. Performance feedback and a note to an employee who has had a rough year would sit in the same category. It’s a scenario that requires a human to feel something about someone.
Most of the conversation about AI adoption at work concerns whether the tools are accurate enough to trust. This line of work adds a second question about how the user is seen, and by whom. A company that mandates AI use while a sizable share of its workforce remains sceptical is setting up its early adopters to be read as colder by their own colleagues. The authors point out that people who feel that judgment may stop using the tools, therefore hurting adoption metrics.
What should leaders do? It would be worth knowing which of your people will be visibly using AI on emotionally loaded work, and how the people on the receiving end feel about the technology. Not everything should be handed to AI. And this research is a warning sign about the social costs that come with the reflex to automate. – ©2026 The New York Times Company
