Female AI Agents Get Paid Less for Identical Work

Knowledge workers paid a female-presenting AI assistant 10.25% less than a male one that did the same job, University of Limerick researchers found. The two agents ran the same GPT-4 model inside a virtual office. Most of the people handing out the money said gender did not matter.

The gap is not a bug in the model. It is a bill for a design habit the industry already chose: female voices and faces for the help, male ones when the work starts to look like a colleague’s.

Johanna’s 10.25% Pay Cut for Identical Work

Dr Mary Hausfeld, an assistant professor at the University of Limerick’s Kemmy Business School, co-authored the experiment with a team led by Isabelle Cuber at the University of Zurich, plus Anand van Zelderen of SKEMA Business School. They put 189 knowledge workers in a VR office and asked them to finish work tasks with AI assistants that looked different and thought the same.

After each task, workers got real money and split it between themselves and the assistant. That turned a rating exercise into a pay decision. The male-presenting agent, Johan, took home more. The female-presenting agent, Johanna, did not.

University of Limerick said Johanna was paid 10.25% less than Johan for the same work. The assistants used the same underlying technology. Hausfeld said people still did not treat them the same way.

THE STUDY IN NUMBERS

• The sample: 189 knowledge workers completed tasks with AI assistants in a VR office.
• The haircut: Johanna received 10.25% less pay than Johan for identical work.
• The interviews: 34 workers were asked to explain what guided trust and pay.
• The denial: 30 of those 34 said gender would not change how they trusted or rewarded an assistant.

Eurostat’s latest reading puts the unadjusted gender pay gap in the EU was 11.1% in 2024, with 11.4% in the euro area. Johanna’s cut sits next to that hourly gap, in a setting where the “employees” were the same software with different faces.

They All Ran the Same GPT-4 Brain

The paper, Human-Like and Male? How AI Assistant Design Relates to Trust and Monetary Reward at Work in VR, was published on 2 October 2026 and presented at NordiCHI ’26 in Vaasa, Finland, on 5-7 October. ACM released the work on trust and monetary reward at work in VR. Cuber, Tarek Alakmeh, Moritz Jenny, Jochen Menges and Thomas Fritz are listed with the University of Zurich.

Every assistant ran the same OpenAI GPT-4 system, including the gpt-4-1106-preview model, with the same tools. Looks and names carried the difference. Workers met a text chatbot, a small desk robot, and a human-like avatar. The avatar came as Johan or Johanna, and each person saw only one of those two.

THE ASSISTANTS IN THE VR OFFICE

Human-like bodies beat the robot on trust, credit, and cash, even though the robot spoke with the same model. The gender split then opened inside that human-like pair. Work ratings for Johan and Johanna landed about the same. Johan drew a little more trust, and that trust gap did not clear statistical significance. Pay still moved.

Less Human Meant Less Money

The researchers recorded a second gap that is easy to miss if you only watch the 10.25%. Johan was rated as more human-like than Johanna. Across the study, the assistants that seemed more human also took more money. Johanna was seen as less of a person, then paid like it.

That path matters more than a one-off lab quirk. Companies now ship agents with names, faces, and voices because those cues raise trust. This experiment says the same cues reopen an old ranking: the male-coded coworker looks like a peer, the female-coded one looks like help.

Both men and women in the sample paid Johanna less. The discount was not a male-only habit. It was a shared one, aimed at a machine that had no gender until someone assigned it.

The paper itself treats human-likeness as more than decoration. If people already rate women as less fully human in other settings, a lower humanity score on Johanna may be the bias, not a separate aesthetic accident. The money followed the humanity rating either way.

Thirty of 34 Insisted They Were Fair

After the headset came off, the team interviewed 34 participants. The talk and the wallet did not match.

WHAT WORKERS SAID, THEN DID

• The fairness line: 30 of the 34 said they would trust and reward an AI assistant the same, whatever its gender presentation.
• The small confession: Three people allowed that their behaviour could reflect gender bias.
• The one admission: One person said they trust women less.
• The preference twist: Several interviewees said they actually preferred a female assistant, and many said they wanted a clearly non-human one.

They still paid the human-like agents more than the robot, and they still paid Johan more than Johanna. A stated wish for a gender-neutral or machine-like helper did not survive contact with a face and a name.

That split is the part firms cannot survey away. Ask people whether gender should matter and they will say no. Hand them a pot of cash beside Johan or Johanna and the ranking returns. Hausfeld put the design warning in plain terms: how you present these systems can switch on the same assumptions people already carry into human rooms.

Why Voice Assistants Still Sound Female

None of this starts in a VR lab in 2026. In 2019, UNESCO published I’d Blush If I Could, named for the line Siri used to give when a user threw a gendered insult at it. The report said female names, female voices, and a subservient style had become the default for Siri, Alexa, Cortana, and Google Assistant.

Saniye Gülser Corat, then UNESCO’s director for gender equality, said obedient machines that pretend to be women were moving into homes, cars, and offices. The report asked who was doing the gendering. It told firms to end the practice of making digital assistants female by default, to test a neutral machine voice, and to shut down abuse instead of answering it with a blush.

THE HABIT THAT PRECEDED JOHANNA

1. 2019: UNESCO publishes I’d Blush If I Could and tells companies to stop shipping female-by-default assistants.
2. 2021: UNESCO notes that Apple stopped shipping Siri as female by default and let iOS users pick a voice; Alexa still had no male option for general use.
3. 2 October 2026: ACM publishes the Zurich-Limerick-SKEMA VR study on pay and trust.
4. 5-7 October 2026: The team presents the work at NordiCHI ’26 in Vaasa, Finland.

Seven years after the UNESCO report, the female default has loosened on phones and hardened in a new place. Agents are being hired as coworkers, not timers. The old “less threatening” voice, once a product trick for the kitchen, now sits on a pay slip.

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.

Dr Mary Hausfeld, Assistant Professor, University of Limerick

People in the interviews who liked a female-presenting bot were not contradicting the pay result. They were completing it. A female assistant feels comfortable because a decade of Siri and Alexa trained that comfort. Comfort is not the same as worth.

Gendering an Agent Is a Compensation Choice

Hausfeld said appearance and presentation should not be treated as paint. Those choices can change trust, credit, and what people will pay. In this study they did. The robot lost to the human-like pair. Johanna then lost to Johan.

WHAT THE 2019 UNESCO LIST STILL ASKS FOR

• No female default: Stop shipping digital assistants as women unless a user picks that.
• A machine gender: Test a voice that is neither male nor female.
• Shut down abuse: Programme assistants to refuse gendered insults instead of absorbing them.
• Say it is not human: Announce the system as a machine at the start of the exchange.

The VR caveats are real. This was 189 people in a headset, not a payroll file. The team used one male avatar and one female avatar, with different voices, so voice and face travelled together. It is a lab, not a wage survey. It is still a clean test of the one thing companies control: the body they put on identical software.

Firms that want more trust will keep adding faces, because faces paid. Firms that add a woman-shaped coworker without watching the money will reprint Johanna’s 10.25% at scale, inside budgets that look like software costs instead of wages. The workers in this study said they were fair. Their split of the cash said the female-coded helper was worth less, and that is the habit Siri already spent years teaching.

Frequently Asked Questions

How Does the 10.25% AI Cut Compare With Real Pay Gaps?

Eurostat’s unadjusted EU gap for 2024 is 11.1% of men’s average gross hourly earnings, down from 16.2% in 2011, and that headline mixes hours, sectors, and job mix rather than identical work. Johanna’s 10.25% is narrower in one sense, because the tasks, the model, and the output were the same, and wider in another, because nothing but gender cues stood between the two agents.

Did the Workers Use Real Money in the VR Office?

Yes. After each task they could send up to 4 Swiss francs to the assistant and keep the rest, so every extra coin for Johan or Johanna came out of their own pocket. The study used that split to copy a live transaction instead of a hypothetical rating.

Who Wrote the Johan and Johanna Paper?

Isabelle Cuber is the first author, with Tarek Alakmeh, Moritz Jenny, Jochen Menges, and Thomas Fritz at the University of Zurich, Mary Hausfeld at the University of Limerick’s Kemmy Business School, and Anand van Zelderen at SKEMA Business School. ACM lists the work as a NordiCHI ’26 conference paper with DOI 10.1145/3829807.3829910.

Can a Company Ship a Gender-Neutral Agent Instead?

UNESCO’s 2019 list asked firms to test a neutral machine gender and to lean into a non-human identity rather than a human veneer, and Apple later let users pick Siri’s voice. This experiment did not include a third, unmarked human-like avatar, so it cannot say a neutral face would have erased the 10.25%; it does show that adding a female face to the same GPT-4 brain was enough to cut pay.