Participants who felt discriminated against chose an AI assessor more often in a campus field study
In a campus field study, participants who felt discriminated against chose an AI assessor more often than those who did not, researchers report. That cuts against the documented preference for human advice, with implications the researchers pose for retailers and service providers.
The Scientist · Science desk
What happened
- Earlier work documented algorithm aversion, a tendency to prefer human advice over AI, and it held even when the algorithm objectively performed better.
- Research confederates posing as agents assessed loan applications on a university campus with near-identical questionnaires, one of which added a question asking for a self-assessed skin color.
- Every applicant was rejected, and those who received the skin-color question reported feeling that they had been discriminated against.
- At a later follow-up, every participant was offered a choice between a human assessor and an AI assessor.
- The researchers attribute the shift to embarrassment: people who feel judged over an irrelevant attribute seek refuge in systems perceived as neutral and free from bias.
Why it matters
- decision The study offered an AI assessor beside a human one, so it bears on adding an AI option next to people. Replacing the human option is a different question the study did not test.
- exposure The pull is toward systems perceived as unbiased. An assistant that in fact treats customer groups differently would be receiving people who came to it looking for refuge.
- constraint If the effect carries into unrelated domains, as the researchers say, a retailer's assistant uptake will mix ordinary preference with the effect of mistreatment elsewhere, and the retailer cannot see which is which.
Because every applicant was rejected [3], rejection itself cannot be what separates the groups. The questionnaires differed in one item, the self-assessment of skin color [2]. That is a clean contrast for a field setting.
The phase 2 comparison is the softer joint. The researchers describe it as a contrast between participants who felt discriminated against and those who did not [5]. That groups people by what they reported feeling. If the analysis splits on the feeling instead of on which questionnaire a person received, part of the gap is correlation, because people who bristle at the question may differ in other ways from people who shrug it off.
The researchers single out discrimination in public places [11]. Embarrassment, in their description, is a social emotion that arises when people feel judged negatively by others [8]. The account as available stops before its experiments are described: it does not report how many people took part, what share chose the AI assessor, or whether embarrassment was measured directly.
The setting is narrow. Participants were on a university campus and chose between assessors of a loan application [2][4]. Carrefour's Hopla assistant appears in the account as an illustration, next to a question about cooking dinner on 20 euros for a guest with a gluten intolerance [7]. The thing this doesn't tell you is how a shopper who has just been mistreated behaves when the stake is a recipe instead of a loan.
In our view the result is a well-motivated hypothesis for operators to test against their own assistant logs: whether customers who report a bad experience pick the assistant more often than those who do not. The researchers call the reversal "algorithm appreciation" [10].
What to watch
- The full paper, with sample sizes, the share choosing the AI assessor, and whether the comparison splits on questionnaire version or on reported feeling.
- A replication in a retail setting with a live assistant and a low-stakes choice, such as a recipe or a product recommendation.
- Evidence on whether the preference survives when the AI assessor is shown to treat groups differently.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption
- Insufficient
- Hype gap+20
- Incentives
- Insufficient
- Confidence40
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
People tend to prefer other people over AI systems, a phenomenon known as algorithm aversion, which has been widely documented even when algorithms objectively demonstrate better performance.
- [2]
In a field study on a university campus, 'agents' who were research confederates assessed loan applications using almost identical questionnaires; one version included an additional question asking participants to self-assess their skin color.
- [3]
All applicants were subsequently rejected; those who had been asked the question about skin color reported feeling that they had been discriminated against (phase 1).
- [4]
In follow-up data collection at a later date, all participants were offered the choice between a human assessor and an AI assessor (phase 2).
- [5]
A significantly higher proportion of those who felt discriminated against in phase 1, compared with those who did not, chose in phase 2 to be assessed by the AI system, perceived to be 'devoid of emotions and biases.'
- [6]
The researchers identify embarrassment as the emotional mechanism: when people feel socially judged or treated unfairly because of an irrelevant personal attribute, they seek refuge in systems perceived as 'neutral and free from bias.'
- [7]
The account uses Hopla, the AI-powered shopping assistant launched by French supermarket chain Carrefour, as an example, alongside a question about what to cook for dinner on a budget of 20 euros for a guest with a gluten intolerance.
- [8]
Embarrassment is a social emotion that arises when individuals feel they are being judged negatively by others.
- [9]
The researchers say consumers who are victims of discrimination are significantly more likely to turn to AI-based recommendations rather than human advice, even in a subsequent domain entirely unrelated to the one in which they experienced discrimination.
- [10]
The researchers call the shift from algorithm aversion to a preference for algorithms 'algorithm appreciation'.
- [11]
The researchers say the effect applies particularly to consumers who are victims of discrimination in public places.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- phys.orgCan AI feel safer than human advice for consumers who face discrimination?
1 article · October 14, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.