Gender Bias in Perception of Human Managers Extends to AI Managers

This study demonstrates that gender biases in leadership perceptions, where male managers are favored and female managers face greater skepticism upon making unfavorable decisions, extend from human leaders to AI managers, highlighting the need to address these prejudices in the design of AI-driven organizational systems.

The Gist

Imagine you are playing a team game with two friends. You all solve puzzles together, and at the end, a "Manager" picks one person to win a small cash prize.

Now, imagine that Manager isn't just a person you know. Sometimes, the Manager is a human. Sometimes, it's an Artificial Intelligence (AI). And sometimes, the Manager is introduced as a man, a woman, or they don't say who they are.

This paper is like a giant experiment to see how we react when that Manager makes a choice. The researchers wanted to know: Does it matter if the Manager is human or a robot? And does it matter if we think the Manager is a man or a woman?

Here is the story of what they found, explained simply:

1. The Setup: The "Robber" Game

The researchers put 556 people online into teams of three. They played a simple game called "Find the Robber," where you have to spot a bad guy in a cartoon picture.

• Round 1: Everyone played alone.
• Round 2: Teams played together.
• The Twist: A Manager (either Human or AI) was assigned to the team. This Manager was randomly labeled as Male, Female, or Unspecified.
• The Prize: The Manager picked one "Best Player" to get an extra 50 pence. (In reality, the pick was random, but the players didn't know that).

Afterward, the players were asked: How much do you trust this manager? Are they smart? Are they fair? Would you work with them again?

2. The Big Surprise: The "Golden Ticket" Effect

The most obvious result was what you'd expect: If you won the prize, you loved the manager. If you lost, you didn't.

• The Winner's Lens: If you got the award, you thought the manager was a genius, super fair, and someone you'd love to work with again.
• The Loser's Lens: If you didn't get the award, you thought the manager was unfair and incompetent.

But here is where the gender bias crept in.

3. The Gender Trap: The "Double Standard"

The researchers found that the type of manager (Human vs. AI) and their gender changed how people reacted, especially when things went wrong.

• The "Male Shield": When a manager was presented as Male (whether human or AI), they were given a lot of grace.

  • If a man won the prize, people thought, "Great choice!"
  • If a man didn't pick you, people were still relatively okay with him. They didn't get as angry. It's like he had a "shield" protecting his reputation.

• The "Female Glass Ceiling": When a manager was presented as Female, the rules were much harsher.

  • Human Women: If a human woman didn't pick you, you were a bit annoyed, but you understood she might have her reasons.
  • AI Women: This was the worst combination. If the manager was an AI presented as a woman and she didn't pick you, people got very angry. They rated her as the least fair, least competent, and least trustworthy.

Think of it like this:
Imagine a male coach and a female coach. If the male coach picks the wrong player, people say, "He's having a bad day." If the female coach picks the wrong player, people say, "She doesn't know what she's doing."
Now, imagine that female coach is a robot. People seem to think, "Not only is she a woman making a mistake, but she's a robot pretending to be a woman making a mistake." The bias against women gets amplified when they are machines.

4. Why Does This Happen?

The paper suggests that we carry our human prejudices into the digital world.

• We Project: We treat AI like people. We give them names, voices, and genders.
• We Stereotype: We have old-school ideas that "Men are leaders" and "Women are helpers." When an AI acts like a leader, we judge it by these old rules.
• The "Uncanny" Doubt: When a female AI makes a decision that feels unfair, we don't just blame the algorithm; we blame the gender of the algorithm. We are quicker to doubt a female AI than a male AI.

5. The Takeaway for the Future

This study is a warning sign for the future of work. As companies start using AI to hire people, give promotions, or manage teams, they might accidentally create a system that is unfair to women.

If an AI manager is designed to sound or look female, and it makes a tough decision (like firing someone or not giving a bonus), that AI might get blamed much more harshly than a male AI or a human man.

The Bottom Line:
Technology isn't neutral. Even though AI is made of code, the people using it bring their biases with them. If we want AI to be fair, we can't just fix the code; we have to fix our own minds. We need to realize that we are judging a robot's gender just as harshly as we judge a human's.

In short: Whether the boss is a human or a robot, if we think the boss is a woman, we are harder on her when she makes a mistake. And if that boss is a robot woman, we are the hardest of all.

Technical Summary

Here is a detailed technical summary of the paper "Gender Bias in Perception of Human Managers Extends to AI Managers" by Cui and Yasseri.

1. Problem Statement

As Artificial Intelligence (AI) transitions from a tool for efficiency to an active agent in organizational decision-making, there is a critical gap in understanding how human employees perceive AI managers compared to human managers. Specifically, it is unclear whether established gender biases in leadership evaluations (where female leaders are often judged more harshly than male leaders) transfer to AI systems.

• Core Question: Does the gender of an AI manager (or the lack thereof) influence perceptions of trustworthiness, competence, fairness, and willingness to collaborate, particularly when the manager makes a decision that results in an award (reward) or non-award (penalty) for an employee?
• Context: The study investigates the intersection of anthropomorphism (assigning human traits to AI), gender stereotypes, and outcome bias (how the result of a decision affects the evaluation of the decision-maker).

2. Methodology

The authors conducted a Randomized Controlled Trial (RCT) using an online experimental platform (Empirica) to simulate a workplace environment.

Participants

• Sample Size: $N = 556$ U.S.-based adults (recruited via Prolific).
• Demographics: Approx. 48.7% female, 48.4% male, 2.9% non-binary/other; average age 37.4.
• Ethics: Approved by University College Dublin (UCD) Human Research Ethics Committee.

Experimental Design

The study employed a mixed factorial design manipulating four primary variables:

1. Manager Type: Human vs. AI.
2. Manager Gender: Male, Female, or Gender-Unspecified.
3. Award Outcome: Whether the participant was selected as the "best player" (Winner) or not (Non-winner).
4. Self-Perceived Contribution: Participants rated their own contribution vs. teammates' contributions to calculate a relative contribution score.

Procedure

1. Pre-treatment Survey: Baseline measurements of trust, competence, fairness, and willingness to work with managers.
2. Task Phase:
   • Round 1 (Individual): Participants solved a visual reasoning puzzle ("Find the Robber") individually.
   • Round 2 (Team): Participants formed teams of three, collaborated via chat, and solved the same puzzle together.
3. Manager Assignment: Participants were randomly assigned a manager (Human/AI) with a specific gender label. They were told the manager would select the "best player" for a £0.50 bonus.
4. Outcome: A random "winner" was announced.
5. Post-treatment Survey: Participants re-evaluated the manager on Trustworthiness, Competence, Fairness, and Willingness to Collaborate (0–10 Likert scales).
6. Contribution Assessment: Participants rated their own and teammates' contributions to derive a "Relative Contribution" metric.

Statistical Analysis

The authors utilized linear regression models with interaction terms to analyze post-treatment perceptions.

• Model: $PF = \beta_0 + \beta_1(\text{Pre-Treatment}) + \beta_2(\text{Demographics}) + \beta_3(\text{Award}) + \beta_4(\text{Contribution}) + \beta_5(\text{Interactions}) + \epsilon$
• Key Interactions: The model specifically tested three-way interactions between Award Outcome × Manager Type × Manager Gender, and Award Outcome × Manager Type × Contribution.

3. Key Contributions

1. Extension of Gender Bias to AI: The study provides empirical evidence that gender bias is not limited to human leaders; it extends to AI decision-makers. The "gendering" of AI (via voice, avatar, or label) triggers the same societal stereotypes as human managers.
2. The "Double Penalty" for Female AI: The research identifies a compounded vulnerability where female AI managers face the harshest negative judgments when they fail to reward an employee, more so than female human managers or male AI managers.
3. Outcome-Dependent Bias: The study demonstrates that bias is dynamic. While initial perceptions show little gender difference, negative outcomes (not receiving an award) trigger significant gender-based disparities in evaluation.
4. Quantification of Relative Contribution: The study introduces a method to quantify "self-perceived contribution" relative to peers, showing that high self-perceived contributors who are not rewarded are significantly more critical of managers, especially human ones.

4. Key Results

• Baseline: Pre-treatment perceptions showed no strong preference based on manager type or gender, though human managers were slightly rated higher than AI.
• Impact of Awards:
  • Winners: Participants who received an award rated their managers significantly higher across all dimensions (trust, competence, fairness).
  • Losers: Participants who did not receive an award rated managers lower.
• Gender Asymmetry in Negative Outcomes:
  • Male Managers (Human & AI): Benefited from positive outcomes and were less penalized for negative outcomes. They maintained higher ratings even when participants did not win.
  • Female Managers (Human & AI): Faced significantly steeper declines in ratings when they did not award a participant.
  • Female AI Managers: Experienced the most severe negative judgments when participants were not awarded, indicating a "double bias" (gender + AI skepticism).
• Demographic Nuances:
  • Female Participants: Showed a stronger negative reaction to being denied an award by a human manager if they perceived themselves as high contributors.
  • Male Participants: Rated male-human managers as less competent and were less willing to work with them again if they were not awarded.
• Interaction Effects: The interaction between Award Outcome, Manager Type, and Manager Gender was statistically significant. The "resilience" of male managers to negative feedback was not observed in female managers.

5. Significance and Implications

• Theoretical Impact: The findings challenge the notion of "technological neutrality." They suggest that social cognition (how humans perceive gender roles) transfers seamlessly to AI, reinforcing the idea that AI systems inherit the biases of their designers and users.
• Organizational Implications:
  • Trust and Morale: As AI takes on supervisory roles, the perceived fairness of these systems will directly impact employee morale. If female AI agents are perceived as less fair or competent during negative outcomes, adoption may be hindered.
  • Design Guidelines: Organizations must move beyond "gender-neutral" as a default. Designers should be aware that assigning gendered traits (voice, name, avatar) to AI can activate harmful stereotypes.
  • Policy: There is a need for "bias-aware" AI interfaces. Strategies might include using gender-neutral avatars or explicitly framing AI decisions as algorithmic rather than agentic to mitigate gendered expectations.
• Future Directions: The authors call for longitudinal studies and cross-cultural research to understand how these biases evolve and how to design interventions (e.g., anonymized AI presentation) to break the cycle of stereotype reinforcement.

Conclusion: The paper concludes that while AI offers potential for objective decision-making, it currently acts as a mirror for human gender biases. Without proactive design and cultural shifts, the integration of AI into leadership roles risks amplifying existing inequalities rather than solving them.