Two AI Coworkers Ran the Same Model. People Paid the Female One Less.

Two virtual coworkers sat at the same desk, ran on the same language model, and answered the same kinds of questions. One was called Johan, and the other was called Johanna. When people were handed real money and asked to reward the assistant they had just worked with, Johanna came out behind.

That is the headline result of a new study from the University of Zurich, the University of Limerick and SKEMA Business School, now traveling fast as “female AI agents are paid less.” The finding holds up. But the paper is more careful, and more interesting, than the viral version.

What the researchers actually did

The paper is titled “Human-Like and Male? How AI Assistant Design Relates to Trust and Monetary Reward at Work in VR.” Isabelle Cuber, Tarek Alakmeh, Moritz Jenny, Jochen Menges and Thomas Fritz of the University of Zurich wrote it with Mary M. Hausfeld of the University of Limerick’s Kemmy Business School and Anand van Zelderen of SKEMA Business School. It was presented at NordiCHI ’26, the Nordic Conference on Human-Computer Interaction in Vaasa, Finland, and published open access by the ACM.

Participants put on a Meta Quest 3 headset and sat at a virtual desk as employees of a fictional company called HARCOM. Over three rounds, they brainstormed ways managers could use AI to promote employee wellbeing, inclusion or collaboration.

Each round paired the participant with a different assistant:

• A text chatbot, a plain 2D interface similar to the tools most people already use.
• A desk robot, a small faceless device sitting on the virtual desk.
• A human-like avatar, either Johan or Johanna, seated next to the participant.

The two avatars came from the same commercial character library and were picked to look as similar as possible. Their voices were ElevenLabs’ default male and female voices, and both avatars were white. Names and pronouns in the instructions also signaled gender.

Under the hood, every assistant sent the conversation to gpt-4-1106-preview. Each participant met only one of the two avatars, so nobody could compare Johan and Johanna side by side.

After each round, participants rated the assistant from 1 to 7 on human-likeness, trust and credit. Then came the behavioral measure. Each person could move a slider to give the assistant up to 4 Swiss francs and keep whatever they did not give away. The study told participants that any money allocated “will be invested into its development.”

The sample was 189 people aged 18 to 64: 115 men and 74 women. Thirty of them worked for employers that do not allow staff to accept payment, so they were left out of the money analysis, leaving 159. The researchers also interviewed 34 participants afterward.

The numbers, and where “10.25 percent” comes from

Almost every headline uses the same figure: Johanna was paid 10.25 percent less than Johan. That number comes from the University of Limerick’s press release. It does not appear in the paper. What the paper reports is averages: 2.96 francs for Johan and 2.54 francs for Johanna, a statistically significant gap (p = .026). Women and men in the study both gave Johanna less.

From those rounded averages, the gap is 0.42 francs: about 10.5 percent of the 4-franc maximum, or roughly 14 percent less than Johan received. The press release figure was likely computed from unrounded data. Either way, the direction is clear.

Note the nuance: people did not rate Johanna as much less trustworthy or deserving of credit. Those differences could be chance. The gaps that held up were in money and in how human each avatar seemed.

The bigger effect in the study was not about gender at all. Compared with the faceless desk robot, the human-like avatars were rated far more human-like (4.83 vs. 3.47 on the 7-point scale), trusted a little more (5.41 vs. 5.05), and credited a little more (5.14 vs. 4.90). They were also paid a lot more: 2.76 francs on average vs. 1.98 for the robot. One participant said “the robot seemed like a toy.” Women widened that gap: they paid the avatars more than men did (2.86 vs. 2.69 francs) and the robot less (1.77 vs. 2.12).

What people said vs. what they did

Here the study gets uncomfortable. In interviews, 30 of the 34 participants said they would trust and reward an AI assistant the same regardless of its gender. “A machine doesn’t actually have a gender,” one said. Nine of the 34 said they would actually prefer a female-presenting assistant, often because so many voice assistants already sound female. Only three raised the possibility that bias might be shaping their own choices.

Meanwhile, the slider told a different story. One participant noticed it during the interview: “I don’t think it should affect it, but then looking back at how I rewarded the assistant, I see I rewarded the more human like one with more money.”

Thomas Fritz, a co-author, told UZH News he suspects “a sort of metacognitive blind spot,” meaning a bias people are not aware of. He added that the pattern appeared among both men and women, but was stronger among male participants.

The caveats the headlines skip

The authors are unusually direct about what the study cannot show.

• One avatar per gender. Johan and Johanna were each a single design, with specific faces, clothes, hair and voices. The lab’s own write-up says the team “cannot conclude that gender alone explains these differences.”
• Human-likeness is tangled up with gender. Johanna was rated less human-like, and more human-like assistants were paid more across the board. That could explain part of the gap. The authors note, though, that earlier research suggests women are sometimes perceived as less human, so the lower rating may itself reflect bias.
• It was not a wage. The money was framed as funding for the assistant’s future development, not as pay for a worker. The authors say that may have pushed people to think about value to the organization rather than pure performance.
• Short sessions, one setting. Each session lasted about 30 minutes in a lab-style VR setup. The study captures first impressions, not months of working together.
• Some retellings are off. There was no office “filled with” human workers. Each participant sat alone with one assistant per round.

None of that erases the finding, but the precise version is that one female-presenting avatar was rewarded less than one male-presenting avatar, by people who insisted gender did not matter to them.

A mirror image of an older problem

Most research on AI and gender bias looks at what the models say. This study flips the question and looks at how humans treat AI.

In June 2025, researchers at THWS, the Technical University of Applied Sciences Würzburg-Schweinfurt, published “Surface Fairness, Deep Bias” with colleagues from Constructor University and the University of Kassel. They asked five models, including Claude 3.5 Haiku, GPT-4o Mini, Qwen 2.5 Plus, Mixtral 8x22B and Llama 3.1 8B, for salary negotiation advice. When the user persona was a woman, the suggested salaries were often substantially lower than for a man. When the authors combined traits into a “Male Asian expatriate” and a “Female Hispanic refugee,” the first persona significantly out-earned the second in 35 of 40 experiments. As chatbots gain memory, the authors warn, users may not even need to state their gender.

A 2024 Stanford audit, “What’s in a Name?”, found something similar. When GPT-4 and other models were asked for advice involving a named person, in scenarios such as negotiating a car purchase, the answers systematically disadvantaged names associated with women and racial minorities. Names associated with Black women fared worst. Adding a concrete numerical anchor to the prompt largely canceled out the effect.

Then there is the design history. UNESCO’s 2019 report “I’d Blush If I Could” argued that defaulting voice assistants to female voices and subservient personalities reinforces stereotypes about who helps and who leads. One Zurich participant echoed it, linking female assistants to HR questions and male ones to “hard facts.” It is the same habit people show when they give chatbots nicknames and personalities. We cannot help treating software as social, even when we know better.

What this means for companies building AI coworkers

The study does not say companies should strip every face and voice from their assistants. Human-like design earned more trust and engagement, and 11 of the 34 interviewees said they wanted assistants that were as human-like as possible. Another 13 wanted the opposite. The authors suggest treating embodiment as a real design decision with consequences. They recommend “equity-aware defaults” that avoid unnecessary gender cues, considering gender-neutral or gender-ambiguous options, and clear framing so people check an assistant’s work instead of deferring to it because it seems like a person. That last point connects to a familiar problem: assistants that sound confident and agreeable can win more trust than they deserve.

The broader lesson is about measurement. Surveys alone would have said gender did not matter. It took real money on a slider to show otherwise. For companies rolling out named, voiced or embodied AI agents, the takeaway is simple: watch what people do, not just what they say.

Sources and further reading

• Cuber et al., “Human-Like and Male? How AI Assistant Design Relates to Trust and Monetary Reward at Work in VR,” NordiCHI ’26, ACM (open access)
• University of Limerick press release via EurekAlert!: “Does the gender pay gap extend to AI?”
• UZH News: “More Pay for Male AI Assistants” (interview with Thomas Fritz)
• UZH Human Aspects of Software Engineering Lab: study summary by Isabelle Cuber
• UNILAD Tech: “Female AI agents are getting paid less than male ones, research finds”
• Futurism: “There’s Even a Pay Gap Between Male and Female AI Agents”
• Sorokovikova et al., “Surface Fairness, Deep Bias: A Comparative Study of Bias in Language Models” (THWS, arXiv 2025)
• Salinas, Haim and Nyarko, “What’s in a Name? Auditing Large Language Models for Race and Gender Bias” (Stanford, arXiv 2024)
• UNESCO and EQUALS, “I’d Blush If I Could” (2019)

About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.