I was in a virtual office when a colleague slid a digital paycheck across the desk—name on the screen: Johan. I watched the same output, the same assistant, treated differently. You felt it then: the future arriving with old prejudices tucked into its code.
Observation: In a controlled VR experiment, participants gave the male-presenting assistant more money.
The University of Limerick recruited 189 knowledge workers and put them in a simulated office with three AI assistants: a text chatbot, a robotic deskbound helper, and a humanlike avatar that appeared either male or female. The assistants performed identically, but the humanlike one earned higher ratings and larger pay allocations from participants.
Most striking: the male avatar, labelled “Johan,” was paid about 10% more on average than the female avatar, “Johanna,” despite identical performance. Participants also rated Johan as more human-like than Johanna, which wired respect and reward to presentation rather than output.
Can AI be biased?
Yes—and not because the code is dramatic, but because people are. The study shows how conscious and unconscious gender bias travels with human cues into AI interactions. When you let voice, name, or a facial animation stand in for identity, the same social scripts that shape hiring and raises in real offices begin to play out around bots.
This matters for companies building assistants: Apple (Siri), Amazon (Alexa), Google Assistant, Microsoft, and even conversational models from OpenAI are not just shipping features; they’re shipping personality and social signals. Choices about voice, name, and gender presentation are design decisions with workplace consequences.
Observation: Participants said they preferred female-presenting assistants but rated them as less human and paid them less.
When asked which presentation they preferred, many people named female-presenting voices—echoing the defaults of Siri and Alexa. That preference coexists with a readiness to value male-presenting agents more when it comes to pay and perceived personhood.
There’s a paradox: companies choose female voices because they read as friendly and approachable, but that same framing can funnel women (and female-presenting tech) into service roles with lower status. The result is like a mirror that flatters men more than women, reflecting social bias back at us through code and interface.
Why do virtual assistants default to female voices?
Design teams and product leaders historically reported that female voices are perceived as more helpful and less threatening. That decision—from Apple’s early Siri choices to Amazon’s Alexa—was driven by user testing and market instincts, not a moral argument. But those instincts have social consequences: the normalization of female servility in voice interfaces reinforces stereotypes and trains users’ expectations.
Observation: Interviews revealed people deny bias even as behavior betrays it.
Researchers interviewed 34 participants. Thirty insisted they’d treat assistants the same regardless of gender presentation; only three admitted their behavior might reflect bias, and one openly confessed, “I trust women less.”
This gap between self-image and action is the core problem: you can believe you are impartial and still react to cues. The office, once physical, becomes a stage where gender scripts are auto-played by the smallest details—name, tone, and face—even when you think you’re being fair.
How should firms respond when AI reflects social bias?
Start by recognizing that design choices are policy choices. Teams at Google, Microsoft, and startups working with OpenAI-style models must treat persona, voice, and naming as variables that affect outcomes. That means running blind evaluations, mixing presentations, and testing compensation decisions in controlled pilots before deploying assistants into workflows that touch pay and performance reviews.
Regulators and HR leaders should also demand transparency about how agent personas are created and audited. If a simulated coworker changes the way you compensate it, it will change how you compensate humans too.
I’ve seen enough experiments to know tech won’t erase old inequalities on its own—so if we let interfaces inherit our biases, who will be left to fix them?