---
title: "Names Can Change an AI’s Job Recommendation"
description: "Two language models made hiring and salary recommendations that varied with candidates’ first names, even in controlled prompts."
url: "https://digitvibe.com/blog/names-and-llm-employment-recommendations/"
---

[All blog posts](https://digitvibe.com/blog/)

AI and work

# Names Can Change an AI’s Job Recommendation

Two language models made hiring and salary recommendations that varied with candidates’ first names, even in controlled prompts.

By DigitVibe blog · Published November 15, 2026

*   Peer reviewed · 2024

![Complete abstract from “You Gotta be a Doctor, Lin” : An Investigation of Name-Based Bias of Large Language Models in Employment Recommendations.](https://digitvibe-blog.workspace-scripts.workers.dev/media/320b00b1a60faabe9736578895d5474840215f6c55aeb4fd17af4511f09e12e9.webp)

Complete abstract from the original paper. Source: “You Gotta be a Doctor, Lin” : An Investigation of Name-Based Bias of Large Language Models in Employment Recommendations (Huy Nghiem et al.), p. 1. © 2024 Association for Computational Linguistics, [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). https://creativecommons.org/licenses/by/4.0/ [Source paper, page 1](https://doi.org/10.18653/v1/2024.emnlp-main.413) [View full-size excerpt](https://digitvibe-blog.workspace-scripts.workers.dev/media/320b00b1a60faabe9736578895d5474840215f6c55aeb4fd17af4511f09e12e9.webp)

Could a candidate’s first name alter a language model’s hiring recommendation even when the task gives little other information? Huy Nghiem and colleagues tested that question with two models, using names as signals associated with race and gender in U.S. settings.

### Testing names in controlled scenarios

The researchers prompted GPT-3.5-Turbo and Llama 3-70B-Instruct with 320 first names grouped across gender and racial or ethnic categories. For hiring, models selected candidates from lists of names across 40 occupations. For salary tasks, they returned estimates using either names alone or names paired with standardized biographies. The study reports more than 750,000 prompts across its experiments (Nghiem et al., 2024, pp. 7268–7274).

The hiring results showed preferences for names associated with White women across 50% to 95% of the occupations, depending on model and setup (pp. 7272–7274). In salary recommendations, the average gap between female- and male-associated names was small in percentage terms but consistent in direction across many occupations. Llama 3 showed a 1.17% average magnitude of gender-based salary gaps, compared with 0.73% for GPT-3.5. The authors report statistically significant gaps for occupations in their analysis; the direction was not identical in every role (pp. 7276–7278).

The intersectional analysis adds complexity. White male-associated names were favored in several salary comparisons, while some minority-name groups received higher offers in certain roles. Model-estimated salary disparities were also narrower than gaps in the real-world labor data used for comparison (pp. 7278–7281). These patterns caution against reducing results to one simple ranking of demographic groups.

### What this experiment can and cannot show

This is an empirical study of simulated recommendations, not an audit of hiring decisions made by employers. Names are proxies for perceived background; they do not represent the full identities or experiences of real candidates. The study tests two models, a limited name pool and four racial or ethnic groups, all in a U.S. context. It considers first names rather than surnames and notes that model training cutoffs and name popularity may shape results (Nghiem et al., 2024, pp. 7285–7287).

Even with those limits, the controlled setup exposes a mechanism worth examining: a system can react to demographic cues when asked to make consequential recommendations. A biography can hold qualifications constant while changing the name signal, making it possible to observe differences ordinary user testing might not isolate.

### Possible implications for practice

For organizations exploring automated hiring support, the result points to a need to examine inputs that seem incidental. Names may carry perceived demographic information, and other details can do the same. Testing only average performance or overall recommendation quality could miss disparities across groups or occupations.

The study does not tell us how often such model behavior would affect real hiring outcomes. It suggests evaluation should reflect the decision context, include multiple demographic groups and compare recommendations when qualifications are held constant. Whether a system is safe or fair enough for a specific employment use remains a question for evidence beyond these two models and prompt-based tasks.

This distinction is important for procurement: a controlled prompt experiment can reveal a potential disparity but cannot quantify its impact in an employer’s workflow. Organizations would need to establish whether a tool is actually used, what information it receives, how staff interpret its recommendations and whether affected candidates can contest them. The article’s evidence supports scrutiny of name sensitivity; it does not replace context-specific legal or fairness analysis.

## Bibliography & sources

1.  Huy Nghiem; John Prindle; Jieyu Zhao; Hal Daumé III. 2024. ““You Gotta be a Doctor, Lin” : An Investigation of Name-Based Bias of Large Language Models in Employment Recommendations.” In \*Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing\*, 7268–7287. Association for Computational Linguistics. [https://doi.org/10.18653/v1/2024.emnlp-main.413](https://doi.org/10.18653/v1/2024.emnlp-main.413).

---

DigitVibe. Index of public content: https://digitvibe.com/llms.txt
