Pay less for an artificial intelligence because it shows a female face. That’s what they did 189 workers during an experiment conducted byUniversity of Limerick together with the University of Zurich and the SKEMA Business School, the results of which ended up in Proceedings of the ACM on Human-Computer Interaction. The assistants involved ran on the same software infrastructure, but the agent was baptized Johanna was seen to be recognised 10.25% less compared to his colleague Johanfor the same amount of work done.
The experimental setup is not limited to a questionnaire. Participants wear a headset, enter a virtual office and complete office tasks supported by three types of assistant: a Purely textual chatbota desk robot and anthropomorphic agents with male or female appearance. In the end they receive real money to be shared between himself and the assistant, a mechanism that transforms the judgment into a measurable economic transaction.
Johan was perceived as more human-like than Johanna. And trust rewards anthropomorphism in general, as agents with bodies and faces have garnered more credit than text-based chatbots and desktop robots.
What is declared and what is done
The contrast lies in the gap between words and behaviors. Many participants said they preferred an overtly non-human AI, and most argued that the assistant’s gender was entirely irrelevant. The evaluations and the sums assigned tell another story, as documented by the research signed by the Limerick group.
Mary Hausfeld of Kemmy Business School emphasizes this very point: “What’s striking about our results is that the technology behind these AI agents was exactly the same, but people didn’t treat them the same way.” On the presumed neutral character of these systems he adds: “We often think of AI as something neutral, but the way we design and present these systems can activate the same assumptions and prejudices.”
The work, entitled Human-Like and Evil? How AI Assistant Design Relates to Trust and Monetary Reward at Work in VRbears the signature of Isabelle Cuber, Tarek Alakmeh, Mary Hausfeld, Anand van Zelderen, Moritz Jenny, Jochen Menges and Thomas Fritz, and was presented at the conference NordiCHI ’26 in Vaasa, Finland, between 5 and 7 October 2026.
The results should be read with caution. The champion of 189 people is small to detect an effect of this magnitude, the two anthropomorphic avatars used are only two and their graphic rendering is not superimposable: objections already circulated in the discussions following publication, which require replications with different faces and more solid numbers to be archived.
However, the operational message also holds up in a prudent form. Choosing the sex, voice and body of a conversational agent is not an aesthetic decision to be entrusted to the marketing department: it affects trust, the merit attributed to the work done and even how much you are willing to pay for it. With AI assistants entering company flows, the real risk is to reconstruct the same asymmetries inside the software that we are trying to dismantle outside.
