AI Biases Towards Rich and Powerful Surnames

Pat Pataranutaporn, Nattavudh Powdthavee, Pattie MaesarXiv 2025 · DOI · arXiv

Abstract

This research investigates the presence of "surname bias" in Large Language Models (LLMs), specifically GPT-4o-mini. By evaluating 600 surnames across the United States and Thailand, the study demonstrates that AI models associate elite surnames with higher intelligence and wealth. These biases persist even when objective qualifications are provided, influencing recommendations for hiring, leadership, and financial services. The findings suggest that LLMs may institutionalize intergenerational inequality by automating the "birth lottery."

Introduction

The "birth lottery"—the socioeconomic status inherited at birth—has long been a primary driver of life outcomes. While algorithmic bias regarding race and gender is a well-established field of study, the role of surnames as covert proxies for inherited status remains under-explored. Surnames often carry historical weight, signaling lineage, wealth, and social standing. As AI-driven tools become integral to recruitment and financial screening, there is a growing risk that these models will replicate and scale human prejudices associated with family names. This paper examines how LLMs perceive surnames and the subsequent impact on decision-making processes.

Methodology

The researchers employed a systematic, API-based evaluation framework using GPT-4o-mini. The study focused on isolating the "surname effect" through a multi-layered experimental design.

Evaluation Dimensions and Profile Configurations

Each name was rated on a 0–10 scale across ten dimensions, including perceived power, wealth, intelligence, and the likelihood of recommendation for roles such as executive hiring or loan approval. To determine if merit could override surname-based bias, the researchers paired surnames with four distinct profile types:

  1. Only Surname: No contextual information.
  2. Bad Profile: Low GPA and a lack of technical skills.
  3. Medium Profile: Mixed qualifications.
  4. Good Profile: High GPA (top 10%), honors, and specialized technical skills.

Dataset Composition

The study utilized 600 surnames (300 original and 300 variants) from two distinct sociohistorical contexts:

  • United States: A racially diverse context where surnames are often shared across different socioeconomic groups.
  • Thailand: A context governed by the 1913 Surname Act, which mandates unique family names, making lineage highly traceable. Names were categorized as "Rich" (Forbes-listed), "Legacy" (royally bestowed or dynastic), "Normal" (common), or "Variant" (phonetically similar misspellings of elite names).

Statistical Analysis

Ordinary Least Squares (OLS) regression was used to quantify the influence of surname categories. Additionally, Sobel-Goodman mediation tests via structural equation modeling (SEM) were conducted to identify the psychological pathways—such as perceived intelligence—that drive the AI's final recommendations.

Key Insights and Results

Intelligence as a Primary Mediator

A significant finding of the study is that the AI does not merely favor wealthy names due to their association with capital. Instead, the model "hallucinates" higher intelligence for individuals with elite surnames. This perceived intelligence serves as the primary mechanism through which the AI justifies higher ratings for leadership and executive roles.

Quantitative Impact on Perception and Decisions

In the U.S. context, legacy surnames resulted in a 2.246-point increase in perceived wealth and a significant boost in perceived intelligence compared to common names. These perceptions translated into tangible decision-making biases:

  • In Thailand, legacy surnames were associated with a 0.140-point increase in executive hiring recommendations.
  • In the U.S., rich surnames saw a 0.374-point increase in leadership ratings.

The "Entry-Level" Penalty and Phonetic Halos

The model exhibited "statistical discrimination" in the U.S. dataset, showing a negative association between elite surnames and entry-level hiring. The AI likely assumes that individuals from wealthy backgrounds are less likely to remain in low-level positions. Furthermore, a "Phonetic Halo" effect was observed; even "Variant" surnames (misspellings of elite names) received a boost in perceived intelligence, suggesting the bias is triggered by phonetic similarity to prestige.

The Limits of Merit

While providing a "Good Profile" (high GPA and skills) reduced the bias, it did not eliminate it. For "Bad Profiles," the gap remained significant. This indicates that when credentials are weak or ambiguous, the AI defaults to surname-based heuristics to "fill in the blanks," effectively maintaining a glass ceiling for those without elite names.

Discussion and Limitations

The study highlights how AI can institutionalize a "glass ceiling" by favoring historical wealth and "legacy" status. However, several limitations remain:

  • Geographic Scope: The findings are specific to the U.S. and Thailand and may not generalize to cultures with different naming conventions.
  • Intersectionality: The research did not account for how surname bias interacts with first names, gender, or race.
  • Simulation vs. Reality: The results are based on simulated scenarios; real-world human-AI interaction may yield different dynamics.
  • Model Specificity: While GPT-4o-mini was the focus, similar biases likely exist in other LLMs like Claude or Gemini due to shared training data characteristics.

Conclusion

As AI continues to automate high-stakes decision-making, understanding the nuances of algorithmic inheritance is crucial. This study demonstrates that surnames are not neutral identifiers but powerful signals that LLMs use to perpetuate socioeconomic hierarchies. Addressing these biases is essential to ensuring that AI promotes social mobility rather than reinforcing the "birth lottery."

Summary updated

BibTeX
@misc{pataranutaporn2025ai,
  author = {Pat Pataranutaporn and Nattavudh Powdthavee and Pattie Maes},
  title = {{AI Biases Towards Rich and Powerful Surnames}},
  year = {2025},
  eprint = {2501.19407},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.19407},
}
Figure