Key Takeaways:
- Female-presenting AI agents received 10.25% less reward
- Human bias can carry into AI interactions
- AI design choices can have workplace consequences
A new workplace study has found that gender bias can influence how people evaluate and financially reward artificial intelligence agents. In a virtual workplace experiment involving 189 knowledge workers, female AI agents received 10.25% less monetary reward than otherwise identical male-presenting assistants for completing the same work.
The finding highlights a growing concern as AI agents increasingly take on roles that resemble digital colleagues, assistants and workplace collaborators. While the technology behind the systems was the same, participants responded differently depending on how the AI was presented.
Researchers from the University of Limerick, the University of Zurich and SKEMA Business School conducted the experiment in a virtual reality office environment. Participants completed work-related tasks alongside different types of AI assistants, including a text-based chatbot, a desk robot and human-like AI agents.
The human-like agents were presented as either male or female and were given the names Johan and Johanna. Both agents had equivalent capabilities and performed the same functions.
After completing the tasks, participants were given real money and asked to allocate a portion of it to the AI assistant based on its contribution. Despite having the same capabilities, the female-presenting assistant received 10.25% less than the male-presenting assistant.
Researchers also found that the male-presenting AI was perceived as more human-like than the female-presenting agent. The finding suggests that design elements such as names, appearance and gender cues can affect how people interpret an AI system and assess its contribution.
The results become particularly relevant as organisations move beyond using AI simply as software and begin deploying AI agents that can communicate with employees, perform tasks and participate in workplace interactions.
People’s Stated Views Did Not Match Their Behaviour
One of the study’s notable findings was the gap between participants’ stated attitudes and their actual behaviour.
Participants generally indicated that the gender of an AI assistant did not matter to them. However, their decisions during the experiment suggested otherwise, with the female-presenting agent receiving a lower financial reward despite performing the same work.
The researchers also observed differences in how participants responded to AI systems based on how human-like they appeared. Human-like AI agents generally received greater trust and recognition for their contributions than the desk robot, even though the systems were powered by the same underlying technology.
The findings indicate that people may carry familiar social assumptions into interactions with artificial agents, even when they consciously believe those factors should not influence their decisions.
This could have implications for how AI agents are designed. Human-like features can make interactions more intuitive and engaging, but they may also introduce social biases that would be less likely to arise when AI is presented as a neutral technological system.
The issue could become more significant as AI agents become more visible in professional settings and increasingly interact with people in roles traditionally associated with human employees.
AI Design Could Become a Workplace Governance Issue
The study does not show that AI systems independently create a gender pay gap. Instead, it demonstrates that people can reproduce gender-based differences when interacting with AI agents that have gendered identities.
That distinction is important as businesses consider how AI agents should be designed and deployed. If employees or customers respond differently to AI based on its perceived gender, seemingly minor design decisions could influence trust, cooperation, evaluation and even financial outcomes.
The findings also raise broader questions about the use of female AI agents in professional environments. An AI assistant given a human name, voice or appearance may be perceived differently from an identical system without those characteristics. Organisations may therefore need to consider behavioural and social effects alongside technical performance when assessing AI systems.
For developers, the research reinforces the importance of examining how users interact with AI rather than evaluating systems solely on accuracy or efficiency. Testing whether different AI personas receive different levels of trust, recognition or reward could become part of responsible AI design.
As AI agents increasingly function as digital coworkers, fairness may depend on more than the algorithms powering them. The way an AI agent looks, sounds and presents its identity can also influence how people value its work.
The study adds a new dimension to the workplace AI debate: even when AI agents have identical capabilities, human perceptions can determine how their contributions are valued.
