Study finds participants rewarded female-presenting AI less

Researchers placed 189 workers in a virtual-reality office and asked them to complete work-related tasks with AI assistants.[1] The study compared three functionally identical assistants with different levels of human-likeness and gender presentation; all three used the same underlying language model.[3] After tasks, participants divided real money between themselves and an assistant, with up to $4 available for each assistant.[1][3] This design let researchers examine how the assistants’ presentation related to rewards while holding their functions and underlying model constant.[3]

Participants rewarded Johanna, the female-presenting assistant, 10.25% less than Johan, the male-presenting assistant, for doing the same work.[1] The male agent was also perceived as more human-like.[1] That result makes the comparison relevant to teams choosing how workplace AI assistants look and present themselves, while cautioning against treating the experiment’s reward difference as a direct estimate of real-world pay disparities.[1]

Across the study, more human-like assistants received higher trust and credit ratings and larger monetary rewards than the robot assistant.[3] Researchers also interviewed 34 participants, and found that people’s stated preferences did not always align with their evaluations and behavior.[3] Many interviewees said they preferred non-human-like assistants and considered gender irrelevant.[3] The contrast suggests that asking users what they prefer may not, by itself, reveal how they will assess or reward an AI assistant in practice.[3]

The researchers said greater human-likeness may promote trust and engagement, but could also create discomfort, miscalibrated expectations or gender stereotypes.[3] HR Dive reported that companies should consider agents’ appearance and presentation as they develop and deploy them.[2] The paper was published on October 2, 2026, in the proceedings of the 14th Nordic Conference on Human-Computer Interaction, and was scheduled for presentation in Vaasa, Finland, from October 5 to 7.[3][4] The evidence points to a practical design question: how might an assistant’s persona influence the judgments users make about its work?[1][3]