arxiv_ai 98% Match Research Paper AI Ethics Researchers,LLM Developers,Policy Makers,Social Scientists,HCI Researchers 1 week ago

Race and Gender in LLM-Generated Personas: A Large-Scale Audit of 41 Occupations

large-language-models › alignment

📄 Abstract

Abstract: Generative AI tools are increasingly used to create portrayals of people in occupations, raising concerns about how race and gender are represented. We conducted a large-scale audit of over 1.5 million occupational personas across 41 U.S. occupations, generated by four large language models with different AI safety commitments and countries of origin (U.S., China, France). Compared with Bureau of Labor Statistics data, we find two recurring patterns: systematic shifts, where some groups are consistently under- or overrepresented, and stereotype exaggeration, where existing demographic skews are amplified. On average, White (--31pp) and Black (--9pp) workers are underrepresented, while Hispanic (+17pp) and Asian (+12pp) workers are overrepresented. These distortions can be extreme: for example, across all four models, Housekeepers are portrayed as nearly 100\% Hispanic, while Black workers are erased from many occupations. For HCI, these findings show provider choice materially changes who is visible, motivating model-specific audits and accountable design practices.

Authors (7)

Ilona van der Linden

Sahana Kumar

Arnav Dixit

Aadi Sudan

Smruthi Danda

David C. Anastasiu

+1 more

Submitted

October 23, 2025

arXiv Category

cs.HC

arXiv PDF

Key Contributions

This large-scale audit of over 1.5 million occupational personas generated by four LLMs reveals systematic underrepresentation (White, Black) and overrepresentation (Hispanic, Asian) compared to BLS data, alongside stereotype exaggeration. It highlights significant racial and gender biases in LLM-generated portrayals across 41 occupations, demonstrating a critical need for improved fairness and alignment in generative AI.

Business Value

Crucial for developers and deployers of generative AI to understand and mitigate potential harms related to bias and fairness, ensuring responsible AI development and preventing reputational damage or discriminatory outcomes.

Paper Metadata

Innovation Type

Empirical Study / Audit

Deployment Feasibility

N/A (this is an analysis paper, not a system to be deployed).

Limitations Addressed

Concerns about racial and gender representation in AI-generated content, specifically occupational personas, and the amplification of societal stereotypes by LLMs.

Performance Gains

N/A (focus is on identifying bias, not performance improvement).

Technical Tags

LLM BiasGender BiasRacial BiasOccupational PortrayalsGenerative AIAI FairnessLarge-Scale AuditPersona GenerationStereotype ExaggerationAI Safety

Research Topics

AI FairnessBias in AILLM EthicsGenerative AI AuditingSocietal Impact of AI

Methods & Architectures

Large-scale auditPersona generationComparative analysis with BLS data Large Language Models (LLMs)

Applications & Tasks

Generative AI Natural Language Generation Societal Impact Studies Systematic under/overrepresentation of demographic groupsAmplification of existing stereotypesUnfair or biased portrayals in AI-generated content Generating occupational personasAuditing LLM outputs for bias

Datasets & Benchmarks

Datasets

Bureau of Labor Statistics (BLS) data

Percentage point difference (pp) in representation

Related Fields

AI EthicsFairness in AINatural Language ProcessingSociologyHuman-Computer Interaction

Keywords

LLM biasfairnessgenderraceoccupationsgenerative AIauditstereotypesrepresentationAI safetypersonalarge language models

Academic Context

#AI Fairness#Bias in AI#LLM Ethics#Generative AI Auditing#Societal Impact of AI

Commercial Potential

Target Industries

TechnologyMediaPublishingEducation

Use Case Examples

Ensuring AI-generated job descriptions are unbiasedDeveloping fair AI companions or virtual assistants

Competitive Edge

Provides a large-scale, empirical analysis of bias in LLM-generated personas, offering concrete data on representation disparities and stereotype amplification, which is essential for guiding future alignment efforts.

Market Opportunity

Growing awareness and market demand for ethical and fair AI solutions.

Revenue Models

N/A

Resource Requirements

Compute Needs

Significant compute required for generating and analyzing 1.5 million personas.

Data Requirements

Access to LLMs capable of generating personas and demographic data (e.g., BLS) for comparison.

Deployment Constraints

N/A

Scalability

N/A

Regulatory Considerations

Findings have implications for AI regulation concerning fairness and non-discrimination.

Production Readiness

Maturity Level

Analysis / Audit

Time to Market

N/A

View Full Paper Back to Papers