Female AI Agents Paid 10% Less: The Surprising AI Gender Bias Study

Female AI agents received 10.25% less monetary reward than male-presenting AI agents in a 2026 workplace study. Researchers found that participants rewarded functionally equivalent AI differently based on gender presentation, highlighting a potential form of AI gender bias as human-like AI agents become digital coworkers.

KumDi.com

A new 2026 study has found that people may reward female-presenting AI agents less than male-presenting AI agents for performing the same work, even when the underlying AI technology is identical. In a controlled virtual-reality workplace experiment involving 189 participants, the female-presenting AI assistant, Johanna, received 10.25% less monetary reward than the male-presenting assistant, Johan.

The finding does not mean that AI systems literally receive salaries or that female AI has a legally defined “pay gap.” Rather, it demonstrates something potentially more important for the future of AI-enabled workplaces: human gender bias can influence how people evaluate, trust and financially reward an AI system based on how that system is presented.

Key Findings at a Glance

Finding	What the study showed
Participants	189 knowledge workers
Environment	Virtual-reality workplace
AI technology	Functionally identical assistants
Male-presenting agent	Johan
Female-presenting agent	Johanna
Monetary difference	Johanna received 10.25% less
Work capability	Essentially the same underlying technology
Human-likeness	Johan was perceived as more human-like
Trust	Human-like agents generally received more trust
Interviews	34 participants were additionally interviewed
Publication	ACM/NordiCHI 2026 conference proceedings

The research was conducted by researchers affiliated with the University of Zurich, University of Limerick and SKEMA Business School and published as Human-Like and Male? How AI Assistant Design Relates to Trust and Monetary Reward at Work in VR.

What Did the Study Actually Find?

The study examined whether the appearance and presentation of an AI assistant could change how humans interacted with it.

Researchers placed 189 participants in a virtual workplace and had them complete work-related tasks with different AI assistants. The assistants included a text-based chatbot, a desk robot and two human-like AI agents.

The two human-like agents were presented as:

• Johan — male-presenting
• Johanna — female-presenting

Importantly, the assistants were based on the same underlying technology and had equivalent capabilities. The researchers therefore had a way to examine whether differences in participants’ behaviour could be associated with the assistants’ presentation rather than a difference in technical performance.

After completing the tasks, participants allocated real money between themselves and the AI assistant as a reward for the work performed.

The result was striking: Johanna received 10.25% less money than Johan.

That difference occurred even though the underlying AI capabilities were the same.

Why Is a 10.25% Difference Significant?

The important point is not that an AI “needs” to be paid.

AI agents do not currently have human employment status, salaries or financial rights in the conventional sense.

The significance lies in human decision-making.

Imagine two AI agents performing the same research task for a company:

Agent A completes the work accurately and is represented as male.
Agent B completes exactly the same work and is represented as female.

If managers, employees or customers consistently perceive Agent A as more competent, trustworthy or valuable solely because of its presentation, the design of the AI can influence economic outcomes.

That becomes increasingly relevant as AI agents move beyond simple chatbots and begin acting as digital coworkers, sales assistants, customer-service representatives, research agents and autonomous workplace systems.

The University of Limerick researchers specifically describe the transition from AI as a tool toward AI as a coworker as an important context for their findings.

How Was the Experiment Conducted?

The research used an immersive virtual-reality workplace rather than a conventional online survey.

This is important because researchers wanted participants to experience the AI assistants in something closer to a workplace interaction.

Step 1: Participants entered a virtual workplace

A total of 189 knowledge workers participated in the experiment. They interacted with AI assistants while completing work-related activities.

Step 2: Participants interacted with different AI designs

The researchers compared AI assistants that varied in human-likeness and gender presentation.

This allowed the researchers to examine whether people responded differently to an AI that looked and behaved more like a human colleague.

Step 3: Participants evaluated the assistants

Participants rated aspects of the AI assistants, including their contribution and characteristics such as trust and human-likeness.

Step 4: Participants allocated real money

Rather than asking participants only, “Which AI did you prefer?”, researchers introduced an economic component.

Participants were given real money and asked to divide it between themselves and the AI assistant based on its contribution.

This created a behavioural measure rather than relying exclusively on stated opinions.

Step 5: Researchers conducted interviews

The study also included interviews with 34 participants, providing qualitative information about how participants understood their interactions with the AI agents.

People Said Gender Did Not Matter — But Their Behaviour Was Different

One of the most interesting findings is the difference between what people said and what they did.

Many participants reported during interviews that gender was irrelevant to their preference for an AI assistant.

Yet their actual reward allocation showed a measurable difference between the male-presenting and female-presenting agents.

This distinction is important in behavioural research.

People may genuinely believe they are evaluating systems objectively. However, subtle design characteristics can still influence decisions without participants consciously identifying those influences.

The University of Limerick researchers highlighted this disconnect between stated preferences and observed behaviour.

In other words:

Conscious preference: “The AI’s gender doesn’t matter.”

Observed behaviour: The male-presenting AI received more money.

This does not prove that every individual participant consciously discriminated against the female-presenting AI. Rather, it suggests that presentation can affect collective human behaviour even when people do not report gender as an important factor.

Did People Trust the Male AI More?

The answer is nuanced.

The researchers found that both human-like AI assistants were generally perceived differently from the less human-like assistants.

The human-like agents received more trust and were given more credit for their contribution than the robot-style alternative. euronews

Within the human-like pair, Johan was perceived as more human-like than Johanna.

However, Euronews reports that while Johan was trusted marginally more, the difference in trust between Johan and Johanna was not statistically significant.

This distinction matters.

It would be inaccurate to say that the research conclusively proved:

“People trust men more than women.”

The stronger and more defensible conclusion is:

The male-presenting and female-presenting AI agents received different treatment, including a significant difference in monetary rewards, despite having the same underlying capabilities.

That is a much more precise interpretation of the evidence.

Why Does AI Appearance Matter?

Traditional software usually does not have a face, body or apparent gender.

A spreadsheet, database or search engine does not necessarily invite users to interpret it as a social actor.

AI agents are different.

Modern AI systems increasingly have:

• Names
• Voices
• Avatars
• Faces
• Personalities
• Conversational styles
• Social roles
• Human-like behaviours

Once an AI is presented as a coworker rather than simply as software, people may begin applying familiar social expectations to it.

This is one reason the study’s title asks:

“Human-Like and Male?”

The research is essentially investigating whether making AI more socially recognizable also makes it vulnerable to the social biases that humans already use when evaluating other people.

The Risk: AI Could Reproduce Existing Workplace Bias

The study does not establish that AI itself possesses gender prejudice.

That distinction is critical.

The underlying AI technology was the same. The observed difference came from how humans interacted with differently presented AI agents.

This creates a potential feedback loop:

Human social bias → AI design → human perception → AI evaluation → economic outcome

For example, suppose an organization designs a customer-service AI with a female voice and another with a male voice.

If customers consistently perceive the male voice as more authoritative, the company might unintentionally assign the male-presenting agent to higher-value interactions.

That could create an additional layer of bias in automated decision-making.

The issue could extend beyond payment to:

• Customer trust
• Sales conversion
• Leadership perception
• Task assignment
• Performance ratings
• Promotion recommendations
• Negotiation outcomes
• Credit for collaborative work
• Customer satisfaction scores

The study itself does not demonstrate all of these outcomes. These are potential applications of the research question, not findings established by this experiment.

Human-Like AI Was Rewarded More

Gender was not the only important result.

The researchers also found that more human-like AI assistants received greater monetary rewards than the less human-like alternatives.

Participants also perceived the human-like agents as more human than the chatbot and robot-style assistants.

This suggests that AI embodiment can influence how people interpret an agent’s contribution.

For companies deploying AI employees or AI coworkers, that creates an important design question:

Should an AI look human at all?

There is no universal answer.

A highly human-like AI may improve engagement and make collaboration feel more natural. But it may also activate social expectations and biases that would not arise with a clearly non-human interface.

The research therefore suggests that AI embodiment should be treated as a behavioural and organizational design decision, not merely an aesthetic choice.

What Does This Mean for Companies Using AI Agents?

Organizations adopting AI agents should not assume that technical neutrality automatically produces behavioural neutrality.

A company may use exactly the same underlying model for two AI agents while users treat them differently because of:

• Voice
• Name
• Avatar
• Gender presentation
• Language style
• Perceived age
• Human-likeness
• Communication style

1. Test AI presentation scientifically

Organizations should A/B test AI interfaces where appropriate.

For example:

• Same model
• Same instructions
• Same task
• Same output
• Different presentation

Then measure:

• Trust
• Task acceptance
• Customer satisfaction
• Conversion
• Reward allocation
• Error tolerance
• Perceived competence

This can reveal whether presentation characteristics are affecting business outcomes.

2. Avoid unnecessary gender coding

If gender does not provide a genuine functional benefit, companies should consider whether assigning strongly gendered characteristics to an AI agent is necessary.

A neutral or clearly artificial presentation may reduce some opportunities for users to project human stereotypes onto the system.

3. Monitor human-AI evaluation

Companies should pay particular attention when humans evaluate AI agents.

For example, if employees are asked to rate AI coworkers, managers should examine whether ratings vary according to the agent’s presentation after controlling for actual performance.

4. Separate performance from personality

AI performance evaluation should rely on measurable criteria wherever possible.

Instead of asking:

“Do you like this AI?”

Organizations should measure:

• Accuracy
• Completion rate
• Response quality
• Revenue generated
• Customer resolution rate
• Time saved
• Error rate

This reduces the risk that subjective impressions dominate performance assessment.

What This Study Does Not Prove

The headline “female AI agents get paid 10% less” is attention-grabbing, but it needs context.

The study does not prove that female AI systems universally receive lower compensation.

It was a controlled experiment involving 189 participants in a virtual-reality workplace.

It also does not establish that:

• All users discriminate against female-presenting AI
• Male AI is inherently more trusted
• Female AI performs worse
• AI systems have legal salaries
• A 10.25% gap exists in every workplace
• The same effect will occur with every AI model
• The effect will remain unchanged as users become more familiar with AI

The researchers themselves studied a particular experimental environment and particular forms of AI embodiment.

Therefore, the most scientifically accurate interpretation is that the experiment provides evidence that gender presentation can influence how humans reward AI agents, not that a universal AI gender-pay gap has been established.

Why This Research Matters as AI Agents Become Digital Coworkers

The significance of this research goes beyond the headline number.

The AI industry is increasingly moving from AI as a tool toward AI as an agent.

A traditional chatbot waits for instructions.

An AI agent can increasingly:

1. Receive a goal.
2. Break the goal into tasks.
3. Search for information.
4. Use software tools.
5. Communicate with people.
6. Make recommendations.
7. Execute actions.
8. Report results.

That makes the AI feel less like software and more like a member of a team.

Once people perceive an AI as a coworker, social psychology becomes increasingly relevant.

The University of Limerick study provides an early experimental indication that the way an AI coworker is embodied can influence human judgement and economic behaviour.

The Bigger Question: Are We Building AI or Rebuilding Human Bias?

The most important lesson may not be about whether AI should be male or female.

It may be about whether organizations understand the difference between technological neutrality and human neutrality.

An algorithm can process the same information regardless of gender.

But the people interacting with that algorithm may not respond to every representation in the same way.

That means AI design has become part of organizational behaviour.

If companies increasingly deploy AI agents as:

• Sales representatives
• Healthcare assistants
• Financial advisors
• Recruiters
• Customer-service agents
• Project managers
• Research assistants
• Administrative workers

then their appearance and personality may have measurable consequences.

The question for AI developers is therefore no longer simply:

“Does the model work?”

It is also:

“How do humans behave toward the system when it enters the workplace?”

FAQs

Bottom Line

The 2026 University of Limerick study provides an important warning for the emerging AI-agent economy: even when AI capabilities are identical, the human presentation of an AI system can influence how people perceive and reward it.

The headline finding is specific: participants in a virtual-reality workplace experiment gave 10.25% less monetary reward to female-presenting Johanna than to male-presenting Johan for the same work.

The deeper finding is more consequential. As AI moves from anonymous software toward human-like digital coworkers, human social biases may become part of the AI interaction itself.

For businesses, the practical lesson is straightforward: AI performance should be evaluated using objective outcomes wherever possible, while the effects of voice, avatar, gender presentation and human-likeness should be tested rather than assumed to be neutral.

This is not evidence that AI itself is sexist. It is evidence that people can bring familiar social biases into their relationships with AI.

And as AI agents become more deeply integrated into the workplace, understanding that human side of AI may become just as important as improving the technology itself.

References

• Euronews (2026). Female AI agents would get paid 10% less than male ones, study finds. Published October 5, 2026. Euronews — Original Article
• EurekAlert! (2026). Does the gender pay gap extend to AI? Research announcement covering the findings of the Nordic HCI study. EurekAlert! — Research Announcement