Alvero and colleagues at Cornell, Pennsylvania and Florida ran a preregistered experiment with 570 U.S. adults on a hiring task. Participants who first saw eight AI-generated attractive faces then rated twelve identical real candidates lower on every measure: hireability by 0.146 points, interview advancement from 77.4% to 72.5%, trust by 0.172 points.
Why this one stopped me
The retouched arm.
The researchers took the same real photos and ran them through a Facetune-style enhancer, calibrated to the exact beauty boost that One-tap Enhance produces (+0.127 SCUT). Those beautified real faces moved nothing downstream. Hireability P = 0.881. Advancement P = 0.657. Trust P = 0.828.
So the drag is not a halo effect from prettier pictures. It is specific to AI-generated people showing up before real ones. The design isolates the mechanism in a way the standard "AI has bias" write-up never manages.
And the direction reverses by stage. When the AI-generated faces were the ones being judged, they rated higher: hireability +0.162, trust +0.130. The synthetic face gets a bump. The real human who comes after pays for it.
Two effects. One image. Only the first is visible.
The design review nobody logs
A head of people at a mid-size B2B SaaS firm is in a 45-minute review for the new careers page. Twelve candidate cards on the deck. Six use AI-generated portraits because the design team wanted to avoid stock-image clichés and could not clear real contractor sign-off by Friday.
The review ends on time. On Tuesday she interviews real candidates for a Customer Success Associate role. The same role, by coincidence, the paper used.
I have sat in dozens of rooms like this one. Pitch decks with generated team avatars. User research with personas built from a prompt. Pricing panels of synthetic customers before a real-user interview round. Everyone in the room knows the faces are synthetic. Everyone also assumes that knowing is enough.
The paper says the knowing is not enough. The retouched null is what proves it. If the problem were general aesthetic bias, beautified real photos would have dragged the same candidates down. They did not, at any P below 0.6.
Study 1 of the paper adds the second concern. In the generator audit of 27 systems, 37% of female-prompted AI-generated faces exceed the median professional fashion model. White-Nigerian attractiveness gaps are positive in all 54 generator-by-measure comparisons at P < 0.001. The default person your model produces is already pre-selected on features your evaluators cannot neutralise by effort.
The downstream drag on real candidates is specific to AI-generated images, not to beautified images generally.
Price the synthetic person into the room before the real one walks in.