Participants Paid Female AI Agent 10% Less Than Identical Male Bot

Participants in a virtual reality experiment paid a female-presenting AI agent 10% less than a functionally identical male-presenting agent — a finding that mirrors the real-world gender pay gap and exposes bias people deny holding.

The study, published earlier this year and charted by researchers examining how humans interact with AI at work, asked participants to collaborate with human-like AI assistants in a virtual office. They then judged the agents on trustworthiness, work quality and deserved pay.

Two agents anchored the experiment: Johan and Johanna. They were identical in functionality. The only difference was gender presentation. Participants ranked the two similarly on credit for their work, with Johan scoring slightly higher on trustworthiness.

What does the pay gap look like for machines?

When participants paid the agents for their work, Johanna received 10% less than Johan on average. Both men and women in the study said the male agent should earn more. Participants also rated the female-presenting agent as significantly less human-like than her male counterpart.

That result clashes with what most participants said in interviews. "If there was an AI man or an AI woman, I would trust the results equally the same," one participant said.

"A machine doesn't actually have a gender and I'm dealing with a machine or with technology," another remarked.

Only three of 34 interviewed participants acknowledged that gender presentation might have influenced their scoring. "I personally think we feel like it doesn't have an impact, but I think it has a very strong impact and so the interpersonal different biases come into play," one of those three said.

The pattern tracks the persistent human wage gap. As of 2025, women in the United States earned 78 cents for every dollar earned by male counterparts, according to the American Progress data cited in the report. Among women working full-time, year-round roles, the gap narrows to 84 cents on the dollar.

How human-like should workplace AI look?

The study didn't stop at Johan versus Johanna. Participants also scored and paid a text-based chatbot and a small robot sitting on their virtual desktop. All four assistants ran on the same underlying large language model. Only the design differed.

Human-like design won every comparison. Participants trusted and credited Johan and Johanna far more than the visibly machine-like agents. They also rated the human-like pair as more human, and paid them more than the chatbot and the robot.

"I gave Johanna the most because, yeah, I could relate the most to her in the way that she was talking and then calling it a 'her' and not 'it,'" one participant said.

"Having Johanna next to me made me more prone to credit her more again because I felt there was a person, a colleague sitting next to me that I had a conversation with and that was giving me her own input," commented another.

For companies deploying AI assistants, the study carries a double implication. Humanized agents earn more user trust — but they also inherit human prejudices, including a pay-valuation gap that users insist they don't hold. Firms racing to anthropomorphize their AI tools may want to test whether the persona they choose changes how customers value the work.