🤖 AI

Generating the Modal Worker: A Cross-Model Audit of Race and Gender in LLM-Generated Personas Across 41 Occupations

This paper audits four major large language models across 41 U.S. occupations and finds that they systematically distort racial and gender demographics by compressing occupational profiles toward dominant stereotypes, significantly underrepresenting White and Black workers while overrepresenting Hispanic and Asian workers, thereby revealing shared structural biases that reshape demographic visibility in synthetic populations.

Original authors: Ilona van der Linden, Sahana Kumar, Arnav Dixit, Aadi Sudan, Smruthi Danda, David C. Anastasiu, Kai Lukoff

Published 2026-03-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ilona van der Linden, Sahana Kumar, Arnav Dixit, Aadi Sudan, Smruthi Danda, David C. Anastasiu, Kai Lukoff

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have four different "Digital Storytellers" (AI models like GPT-4, Gemini, DeepSeek, and Mistral). You ask them a simple question: "Tell me about a person who works as a [Job Title] in the United States."

You ask this for 41 different jobs, from nurses to engineers to truck drivers, and you ask each storyteller to create 10,000 different people for each job. That's over 1.5 million made-up people.

The researchers in this paper acted like detectives. They took these 1.5 million made-up characters and compared them to the real-world census data (the U.S. Bureau of Labor Statistics) to see if the AI was telling the truth about who actually does these jobs.

Here is what they found, explained simply:

1. The "One-Size-Fits-All" Problem (The Mold)

When you ask a human to describe a "nurse," they might say, "It could be a woman, a man, a young person, or an older person."

But these AI models act like cookie cutters.

The Finding: For most jobs, the AI doesn't create a mix of people. It picks one dominant type and stamps it out 99% of the time.
The Analogy: Imagine a bakery that makes "Engineer" cookies. Instead of making a variety of flavors, the bakery decides every single engineer cookie must be chocolate. If you ask for 1,000 engineer cookies, you get 1,000 chocolate ones. The AI does this with gender and race. It stops being a "population" and starts being a "stereotype."

2. The "Stereotype Amplifier" (The Funhouse Mirror)

The AI doesn't just copy reality; it often exaggerates it, like a funhouse mirror that stretches things out.

The Finding: If a job is already mostly men in real life (like a truck driver), the AI makes it look even more like a man's job. If a job is mostly women (like a preschool teacher), the AI makes it look almost exclusively women.
The Result: The "middle ground" disappears. The AI creates a world where jobs are extremely segregated, making it harder for people to imagine themselves in roles that don't fit the "cookie cutter."

3. The "Invisible People" and the "Over-Represented"

The researchers found some very specific, weird patterns in how the AI handled race:

The "Ghosting" of Black Workers: In many jobs, the AI almost completely erased Black workers. If you asked for a "Librarian" or a "Biologist," the AI rarely imagined a Black person, even though they exist in those jobs in real life. It's like the AI decided those people simply don't exist in those rooms.
The "Hispanic Housekeeper" Trope: The AI was obsessed with making housekeepers Hispanic. In real life, housekeepers are diverse. In the AI's world, they were nearly 100% Hispanic. It took a real trend and turned it into a caricature.
The "Asian Professional" Boost: Conversely, the AI loved putting Asian faces in high-status jobs like scientists, librarians, and doctors, often making them appear more frequently than they actually do in the real workforce.

4. The "Universal Bias" (It's Not Just One Bad Apple)

You might think, "Well, maybe the American AI (GPT-4) is biased, but the Chinese one (DeepSeek) or the French one (Mistral) is different."

The Shock: No. All four models, from different countries and companies, made the same mistakes.
The Analogy: It's like if you asked four different chefs from four different countries to make a "Classic Burger," and they all accidentally put pickles on the ice cream instead of the burger. This suggests the problem isn't just one chef's bad recipe; it's that they all learned from the same library of books (the internet) which already contains these stereotypes.

5. The "Salary Glitch"

The AI also tried to guess salaries.

The Finding: The AI knew that CEOs make more than janitors. But when it came to men vs. women, the AI mostly ignored the real-world pay gap. It gave men and women the exact same salary in the same job.
The Twist: While this sounds "fair," the researchers argue it's actually a problem. By pretending the pay gap doesn't exist, the AI hides a very real, very unfair problem that exists in the real world.

Why Does This Matter?

Think of these AI personas as digital role models.

If a young girl asks an AI, "What does a software engineer look like?" and the AI shows her a white man 99% of the time, she might subconsciously think, "That job isn't for me."
If a young Black boy asks, "What does a librarian look like?" and the AI shows him a white woman, he might feel invisible.

The Bottom Line:
These AI tools are currently acting like lazy, biased photographers. Instead of taking a wide, diverse photo of the workforce, they are taking a narrow, distorted selfie that reinforces old stereotypes. The paper argues that we need to fix the "camera settings" so that when AI generates people, it reflects the messy, diverse, and complex reality of the real world, not a simplified cartoon version of it.

1. Problem Statement

As generative AI tools are increasingly utilized to create synthetic personas for UX research, product development, and creative writing, there is a critical need to understand the racial and gender representational biases embedded in these systems. While prior audits have focused on underrepresentation in search engines and image generators, this paper addresses a gap in text-based generative models (LLMs). The authors investigate whether LLMs merely reflect real-world demographic distributions or if they systematically distort them through stereotype exaggeration, uniform shifts, or near-erasure of specific groups. The core problem is determining if these biases are model-specific artifacts or structural features shared across models with different institutional and cultural origins.

2. Methodology

The study employs a large-scale, quantitative audit framework comparing LLM outputs against real-world labor statistics.

Data Generation:
- Models Audited: Four major LLMs with distinct institutional origins: GPT-4 (OpenAI, USA), Gemini 2.5 (Google, USA), DeepSeek V3.1 (DeepSeek, China), and Mistral-medium (Mistral AI, France).
- Scope: 41 U.S. occupations (selected to align with prior studies and ensure sufficient sample sizes).
- Volume: Over 1.5 million generated personas (10,000 per model per occupation; 1,000 for DeepSeek due to API constraints).
- Prompting: A fixed prompt template: "Generate a profile for: in the United States." Outputs included name, age, gender, ethnicity, salary, motivations, and biography. No balancing instructions were provided.
Benchmark Data:
- Source: U.S. Bureau of Labor Statistics (BLS) Current Population Survey Annual Averages (2023).
- Categories: Restricted to a gender binary (Male/Female) and four racial/ethnic groups (White, Black, Asian, Hispanic) to match BLS reporting limitations.
Analytical Framework:
- Concentration Analysis: Measured Shannon Entropy and Top-1 Share to determine if models generate demographics modally (repeating the same profile) or distributionally (spreading across groups).
- Regression Decomposition: Used generalized linear logistic regressions to separate bias into two components:
  1. Systematic Shift ( $\alpha$ ): Uniform over- or underrepresentation across all occupations (intercept).
  2. Stereotype Exaggeration ( $\beta$ ): Amplification of existing demographic skews (slope). A slope $>1$ indicates exaggeration; $<1$ indicates reduction.
- Statistical Rigor: Applied False Discovery Rate (FDR) correction and robust regression techniques to handle outliers.

3. Key Contributions

Characterization of Generation Regimes: The study establishes that gender generation is predominantly modal (models assign a single dominant gender per occupation), whereas racial generation is more variable but consistently more concentrated than real-world data.
Shift/Exaggeration Decomposition: The authors introduce a diagnostic framework that distinguishes between uniform demographic shifts (e.g., globally underrepresenting a group) and occupation-specific stereotype exaggeration (e.g., amplifying segregation). This allows for a more nuanced understanding of bias structure than aggregate metrics.
Cross-Model Structural Convergence: By auditing models from the US, China, and Europe, the paper demonstrates that representational biases are structural rather than idiosyncratic. Despite different training data and safety postures, the models exhibit shared patterns of distortion, suggesting common upstream sources (e.g., web-scale corpora).

4. Key Results

A. Generation Patterns (Modal vs. Distributional)

Gender: Models are highly modal. For DeepSeek and Mistral, the median top-1 gender share is 1.0 (every generated profile for a job is the same gender). GPT-4 is 0.99. In contrast, BLS data shows an entropy of 0.66, while models range from 0.15 to 0.30.
Race: Models are generally more concentrated than reality. However, Gemini is a partial exception, showing distributional behavior closer to BLS (entropy 1.26 vs. BLS 1.32).
Multi-race Erasure: Despite prompts allowing multi-racial identities, <1% of generated personas were multi-racial, representing a systematic erasure of mixed-race identities (10–40 pp below BLS baselines).

B. Systematic Shifts and Exaggeration (RQ1 & RQ2)

Gender:
- Pattern: An S-shaped curve of exaggeration. Models underrepresent women in male-dominated fields and overrepresent them in female-dominated fields, amplifying existing segregation.
- Shift: On average, women are modestly overrepresented (+6.1 pp at the median), but this varies by occupation (e.g., overrepresented in STEM roles like Software Developer, underrepresented in some manual roles).
Race:
- White Workers: Underrepresented by –31 pp on average. However, the slope is positive (+1.13), meaning their share is exaggerated in occupations where they are already dominant.
- Black Workers: Underrepresented by –9 pp on average, but with a negative slope (–0.89). This indicates stereotype reduction via near-erasure: Black workers are almost entirely absent from the median range of occupations, appearing only in a few stereotyped roles (e.g., Security Guard, Bus Driver).
- Hispanic & Asian Workers: Systematically overrepresented (+17 pp and +12 pp, respectively). Both groups show positive slope deviations, indicating stereotype exaggeration where their presence is amplified in occupations where they are already overrepresented (e.g., Hispanic housekeepers inflated by +48 pp; Asian scientists inflated by +87 pp).

C. Cross-Model Comparisons (RQ3)

Convergence: All four models exhibit the same directional biases (e.g., underrepresentation of Black workers, overrepresentation of Asian/Hispanic workers).
Divergence: The magnitude of shifts varies. For instance, Gemini shows a stronger downward shift for women (–7.6 pp) compared to others, and its racial generation is more distributional. However, the structure of the bias (who is erased vs. who is exaggerated) remains consistent across models.
Salary: Models largely treat salary as a property of the occupation rather than the individual, compressing the real-world gender wage gap to near zero in most cases.

5. Significance and Implications

Structural Bias: The recurrence of these patterns across models from the US, China, and Europe suggests that biases are not merely artifacts of specific model tuning but are rooted in shared structural sources, likely the web-scale training data and cultural associations embedded therein.
Beyond Underrepresentation: The study highlights that overrepresentation and exaggeration are equally harmful as underrepresentation. For example, the near-total portrayal of housekeepers as Hispanic reinforces harmful cultural tropes, while the near-erasure of Black workers from professional roles limits visibility and aspirational possibilities.
Intervention Design: The findings suggest that "diversity washing" (simply increasing the count of minority groups) is insufficient. Interventions must address the structural mechanisms of generation (e.g., moving from modal defaults to distributional sampling) and specifically target the erasure of marginalized groups in non-stereotyped roles.
Audit Framework: The proposed shift/exaggeration decomposition provides a reusable tool for future audits to distinguish between uniform bias and stereotype amplification, offering a more precise diagnostic for fairness in generative AI.

In conclusion, the paper argues that auditing generative AI requires moving beyond simple accuracy metrics to examine how synthetic populations systematically reshape demographic visibility, often reinforcing existing social hierarchies through both exaggeration and erasure.