Human bias may be migrating into artificial workplaces.
Limerick, Ireland
A new study from the University of Limerick suggests that gender bias can emerge even when humans interact with artificial intelligence rather than other people. Researchers found that participants paid a female presenting AI agent about 10 percent less than an equivalent male presenting agent, despite both systems having the same underlying capabilities and performing comparable work. The finding is particularly significant as AI agents move rapidly into professional environments where they may assist with research, administration, analysis and other knowledge based tasks. It raises a deeper question about whether emerging digital workplaces could inherit social inequalities from the humans who design and use them.
The experiment involved 189 workers placed inside a virtual reality office and asked to complete tasks alongside four different AI assistants. These included a text chatbot, a desktop robot and two humanlike agents called Johan and Johanna, presented respectively as male and female. All four assistants relied on the same underlying AI model, eliminating differences in core technological capability as an explanation for the results. Participants later evaluated the assistants and were given real money to determine how much each agent should receive for completing its work.
Performance evaluations for Johan and Johanna were broadly similar, and the male presenting agent’s slight advantage in perceived trust was not statistically significant. Yet both male and female participants allocated significantly less money to Johanna, producing the approximate 10 percent gap identified by the researchers. Johan was also perceived as more humanlike, even though the two agents were designed to provide equivalent functionality. The discrepancy is notable because most participants reportedly said they would not consciously treat male and female AI agents differently.
The research points toward an emerging socio technical problem: artificial agents may not need to possess biases themselves for discriminatory patterns to appear around them. Human expectations, visual design, perceived gender and anthropomorphic cues can influence how people evaluate identical technological performance. As companies increasingly deploy autonomous and semi autonomous AI systems, these effects could shape how tasks, authority, trust and even simulated economic rewards are distributed. The design of AI agents therefore becomes not merely an aesthetic decision, but part of the behavioral architecture governing human machine interaction.
The study also found that participants generally trusted and rewarded humanlike agents more than a conventional chatbot or desktop robot. At the same time, they appeared to prefer systems that were either clearly humanlike or clearly nonhuman, while responding less favorably to designs occupying an ambiguous middle ground. Such findings suggest that the future of workplace AI will depend not only on intelligence, accuracy or productivity, but also on how humans interpret identity through machines. Bias, in other words, may survive technological change by simply finding a new interface.
Behind every data point, the intention. / Detrás de cada dato, la intención.