Limerick study participants paid a female-presenting AI assistant 10.25% less than its male counterpart

For an operator, the useful detail is where the gap showed up. Participants did not trust or credit Johanna significantly less than Johan. The differences came in how human-like she seemed and in money ^[5]. For every 100 units of reward Johan received, Johanna received about 89.75 ^[14]. The researchers wrote: "While participants articulated fairness-oriented principles and design preferences in interviews, their evaluations and allocation decisions in practice varied with human-likeness and gender cues" ^[9].

That split matters because rollout teams usually collect what staff say they think. On this evidence, a survey asking staff whether they trust an agent would have scored the two assistants as equals ^[5]. Dr Mary Hausfeld, a co-author of the report, said: "What is striking about our findings is that the technology behind these AI agents was exactly the same, but people did not treat them in the same way" ^[10].

A skeptic would say nobody pays a software agent, so a reward game tells a manager little. The objection is partly fair. The experiment had 189 participants hand out monetary rewards after working with the assistants ^[2]. In a workplace, software gets valued by how often staff use it and whether it survives a renewal review, and the study did not measure either. Still, the reward shows directly how much value people put on identical work ^[1].

Choosing a persona costs something whichever way it goes. Johan and Johanna both drew significantly higher rewards than the desk robot, which ran on the same underlying model ^{[3] [4]}. The human-like presentation raised what people were willing to give. The female-presenting version then lost part of that premium, and both women and men participants paid her less ^[7]. The study's authors wrote that "embodiment choices for AI coworkers are not merely aesthetic but can meaningfully influence trust, attribution, and economically consequential decisions" ^[12].

Adoption is already broad enough to make this a question for this year. The HP Work Relationship Index 2026 found that 47% of all workers now use AI agents as part of their work ^[11]. Suppose a team picks a named, gendered persona this quarter and judges the agent on staff feedback next quarter. It may then be reading feedback that the persona helped shape. HCAMag set the 10.25% result beside a global human pay gap of around 18% ^{[6] [15]}. The two figures measure different things, but Hausfeld's worry applies to both: "We don't want to inadvertently reproduce existing inequalities in a new technological setting," she said ^[13].