Workers awarded female-presenting AI 10.25% less money in workplace study

Giving AI a human appearance did not make workplace judgments more equal. In a virtual-reality study of 189 workers, participants awarded a female-presenting assistant 10.25% less money than a male-presenting counterpart for the same work. Both used the same underlying technology, according to a University of Limerick account, pointing to bias in how people valued their contributions.

A virtual office, with real money

The researchers set out to examine whether an AI assistant’s presentation changes how people treat it at work. The team included researchers from the University of Zurich, the University of Limerick and SKEMA Business School. Workers entered a virtual office and completed work-related tasks alongside assistants with different forms: a text chatbot, a desk robot and humanlike agents.

The humanlike assistants included male-presenting Johan and female-presenting Johanna. Their underlying capabilities were the same. After completing the tasks, participants received real money to divide between themselves and the assistant. That made the reward decision more than a question about which avatar people liked: participants had to allocate money in recognition of the work done.

Presentation changed the reward

Johanna received 10.25% less money than Johan for completing identical work. Participants also perceived Johan as more humanlike. The researchers interpret the difference in rewards as evidence that familiar workplace gender biases could reappear in interactions with AI, even without a difference in the technology or its capabilities.

Gender was only one part of the experiment. Compared with the desk robot, humanlike assistants attracted more trust and received more credit for their contributions. Because the systems shared the same underlying AI, the findings concern how workers responded to presentation—not a comparison showing that one assistant’s technology performed better than another’s.

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.

Mary Hausfeld, University of Limerick’s Kemmy Business School

Preferences did not predict behavior

Many participants said they preferred AI that was clearly nonhuman. Most said an assistant’s gender did not matter to them. Their evaluations and reward decisions nevertheless showed that these characteristics could influence their behavior. The mismatch is a separate finding from the payment gap: people’s stated preferences did not reliably describe how they treated the assistants.

Hausfeld argues that appearance should therefore not be treated as a purely aesthetic design decision. She says those choices can affect trust, judgments about an assistant’s contribution and financial rewards. Her warning focuses on the characteristics designers give AI and the behavior those choices may encourage in users.

The measured gap concerns reward allocations in a virtual-reality experiment, not salaries paid by employers. It also concerns people judging AI, rather than AI deciding what human workers should earn. The researchers’ warning is about reproducing existing inequalities through interface choices; the experiment does not demonstrate that a particular alternative design would eliminate them.