Identical AI Assistants, Unequal Pay: Male Avatar Given More Cash, Study Finds

Workers Said AI Gender Wouldn’t Matter, but a Male Avatar Earned More

Two virtual office assistants ran on the same AI model, with the same capabilities. Only their looks, voices, and names differed, one built to seem male and the other female. When 189 workers in a virtual reality study split real money between themselves and the assistants’ development, the male-presenting one came out ahead.

Johan, the male avatar (a digital human character), was given an average of 2.96 of a possible 4 Swiss francs, compared with 2.54 for Johanna, the female avatar. Participants also rated him as more human-like. Both men and women gave Johan more. Workers did not clearly trust or credit him more, so money and human-likeness ratings showed the clear gaps.

Afterward, 30 of 34 interviewed participants said an assistant’s gender would not change how they trusted or rewarded it. Yet across the whole group, the money went the other way. The findings appear in the Proceedings of the 14th Nordic Conference on Human-Computer Interaction (NordiCHI ’26).

Workers Paid Male AI Assistant More in Virtual Office Test

Researchers at the University of Zurich, the University of Limerick, and SKEMA Business School recruited 189 knowledge workers, people in jobs built around information and ideas. Recruitment ran through professional networks, university events, and local company ads. Participants were 18 to 64 years old, averaging 35, and 115 were men and 74 were women.

Each wore a headset and sat at a virtual desk, playing an employee of a fictional company called HARCOM. Across three rounds, they worked with three assistants: a plain text chatbot, a small faceless desk robot, and a speaking, human-like virtual avatar. About half were paired with a male avatar named Johan, and the rest with a female avatar named Johanna.

In each round, participants worked with the assistant to come up with three solutions to a workplace challenge about employee wellbeing, inclusion, or collaboration. They then rated the assistant on human-likeness, trust, and credit, each on a scale of 1 to 7, and divided up to 4 Swiss francs between themselves and the assistant’s future development. Whatever they did not allocate, they kept. Because some employers barred workers from accepting payment, the money analysis covers 159 people. The team also interviewed 34 randomly chosen participants. A pilot with 28 people had found the two avatars looked similarly human-like and more so than the robot.

Human-Like AI Assistants Earned More Trust and Money Than a Robot

Johan and Johanna together beat the faceless robot on trust, credit, and money, with money showing the clearest gap: an average of 2.76 francs versus 1.98. Trust and credit gaps were smaller but still solid. Participants were blunt about the robot. One said, “the robot seemed like a toy and it didn’t feel confident in the answers.” Others compared the avatars to “a work colleague,” and some credited them with understanding questions better, though every assistant ran on the same model. One who gave Johanna the most said, “I could relate the most to her.”

Women trusted the assistants more than men did overall, and the money gap between the avatars and the robot was wider among women.

Trust in the plain text chatbot, rated least human-like of all, did not differ meaningfully from trust in the others, and two participants credited it for answers they could copy straight into their work. Some participants called wearing a headset all day impractical, saying chat fit their routines better.

Preferences split. Thirteen of the 34 interviewees wanted clearly artificial assistants, with one saying, “I don’t want to forget that it’s a machine.” Eleven others wanted assistants as human as possible, valuing a colleague to talk to when working alone. Many also disliked designs stuck in the middle, human-like but not fully human, calling them uncomfortable.

Interviewees Said AI Assistant Gender Did Not Matter, but Rewards Differed

One interviewee put it this way: “a machine doesn’t actually have a gender.” Nine of the 34 interviewees said they would prefer a female-presenting assistant, some citing familiarity with the many female AI voices. A few linked female assistants to people-focused questions and male ones to technical facts, a familiar workplace stereotype. One said, “when it comes to hard facts, a man.”

Only three of 34 considered that bias might be at work, and one said, “I think it has a very strong impact.” Looking back at how the money was split, a different participant said, “I see I rewarded the more human like one with more money.”

Pinning the gap on gender alone is not possible. Johan and Johanna were each a single design, with specific clothing, hair, and voices, and the robot, rated even less human-like, also drew less money, so human-likeness alone could explain part of the gap. The authors add that prior research has found women sometimes perceived as less human, so the lower rating itself could reflect bias.

Whatever drove the gap, companies giving AI assistants names, faces, and voices should test how people actually treat them, because asking workers alone would have missed it.

---