Study: Workers Paid Virtual AI Agent 10.25% Less When It Looked Female

The setup: real money, virtual coworkers

Researchers at the University of Limerick, working with colleagues from the University of Zurich and SKEMA Business School, put 189 human participants into a virtual reality office and had them complete work tasks alongside four kinds of AI assistants, according to the Irish Times. The options were a text-based chatbot, a desk robot, and two human-like avatars: a male-presenting one named Johan and a female-presenting one named Johanna.

All four ran on the exact same underlying technology, OpenAI's GPT-4-1106-preview, according to Euronews and IndexBox. Same model, same capabilities, same output. The only variable was how the assistant looked.

After the tasks, participants got actual cash and had to decide how much to keep for themselves and how much to give the AI that helped them, simulating a real payment decision rather than a survey answer.

The number: 10.25%

Johanna was paid 10.25% less than Johan for the identical work, according to the Irish Times, BioEngineer.org, Gadget Review and Moneycontrol. Euronews and IndexBox round that to roughly 10%. Both men and women in the study underpaid Johanna, according to Euronews.

Johan was also rated as more human-like than Johanna, per the Irish Times and Moneycontrol, even though nothing about the underlying system differed.

The study, titled "Human-Like and Male? How AI Assistant Design Relates to Trust and Monetary Reward at Work in VR," is set to be presented at the 14th Nordic Conference on Human-Computer Interaction in Vaasa, Finland, and published in Association for Computing Machinery conference proceedings, according to BioEngineer.org. It's co-authored by Dr. Mary Hausfeld of UL's Kemmy Business School, along with Isabelle Cuber, Tarek Alakmeh, Moritz Jenny, Jochen Menges and Thomas Fritz of the University of Zurich, and Anand van Zelderen of SKEMA Business School.

People say one thing, pay another

Participants told researchers in interviews that gender presentation didn't affect how they treated the AI agents, according to Euronews and Moneycontrol. Their actual payment behavior contradicted that.

"We often think of AI as being neutral, but the way we design and present these systems can activate those same assumptions and biases that exist in our interactions with other people," Hausfeld said, according to the Irish Times and Gadget Review.

People didn't just claim neutrality. Their actions revealed an unstated preference that favored the male-coded agent.

The caveat worth taking seriously

Gadget Review flagged something the other coverage glossed over: the researchers themselves don't claim this proves every female-presenting AI will be shortchanged in every context. "The available evidence does not establish that each feature operates independently or that every female-presenting AI assistant will be treated unequally in all contexts," the outlet noted, citing the study's own framing.

This is one lab experiment with 189 people in a VR headset deciding how to split a relatively small pot of money, not a workplace audit of real corporate AI deployments. The dollar amounts involved were modest, the setting was artificial, and what people do with found money in a lab isn't identical to how a company actually prices AI-agent licensing or compensates human employees. The finding is real and measured, but it describes a human psychological tendency triggered by appearance cues, not a verified pattern in commercial AI pricing.

A separate, related data point

A different story adds context without being the same study. Valerie Chapman, founder of a startup called Ruth AI, told CBS News in September that large language models can give worse salary-negotiation advice to users who present as women. She cited a 2025 study out of Berlin in which an LLM told a male job-seeker to ask for $400,000 for a doctor position, and told an identically qualified female candidate to ask for $280,000, according to CBS News.

Dr. Nici Sweaney, CEO of AI Her Way, told CBS News she tested this herself. Claude initially suggested she price a consulting package at $85,000 framed as "a collaboration," then suggested $150,000 to $250,000 framed around "transformational change" once she told the chatbot she was actually a 52-year-old white man.

That's a different mechanism—LLM text output trained on historical pay data, rather than the Limerick study's finding about human behavior toward a visual avatar. Both point the same direction: AI systems trained on or interacting with human assumptions don't automatically erase the biases baked into those assumptions.

What's unresolved

The Limerick team hasn't published which specific visual or vocal cues drove the gap, whether it holds with larger sample sizes outside a VR lab, or whether it persists when real corporate budgets, not a one-time cash split, are on the line. Until a larger or commercial-scale study replicates the result, this remains a documented lab finding about human behavior, not proof of how AI agents get priced in the actual economy.