# Are Emily and Greg Still More Employable than Lakisha and Jamal? Investigating Algorithmic Hiring Bias in the Era of ChatGPT

**Authors:** Akshaj Kumar Veldanda, Fabian Grob, Shailja Thakur, Hammond Pearce, Benjamin Tan, Ramesh Karri, Siddharth Garg  
**arXiv:** [2310.05135](https://arxiv.org/abs/2310.05135)  
**Date:** 2023-10-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.05135v1/

## Key points

- Study focuses on the use of Large Language Models (LLMs) in algorithmic hiring
- Two tasks examined: matching resumes with job categories and summarizing resumes with employment relevant information
- Potential biases related to protected attributes like gender, race, maternity status, pregnancy status, and political affiliation are investigated
- LLMs demonstrate robustness across race and gender in matching resumes with job categories
- Variations in performance observed for pregnancy status and political affiliation
- Contrastive input decoding technique used to uncover potential sources of bias within open-source LLMs
- Research contributes to understanding algorithmic hiring bias in the context of LLMs
- Highlights strengths and limitations of LLMs in matching resumes while considering protected attributes
- Findings provide insights for improving fairness and mitigating biases in algorithmic hiring processes.

## Summary

The study titled "Are Emily and Greg Still More Employable than Lakisha and Jamal? Investigating Algorithmic Hiring Bias in the Era of ChatGPT" delves into the use of Large Language Models (LLMs) such as GPT-3.5, Bard, Claude, and Llama in algorithmic hiring. The researchers focus on two tasks: matching resumes with job categories and summarizing resumes with employment relevant information. They also examine potential biases related to protected attributes like gender, race, maternity status, pregnancy status, and political affiliation. By replicating a seminal experiment conducted by Bertrand & Mullainathan in 2003 and applying it to state-of-the-art LLMs, the researchers aim to evaluate bias or lack thereof across various attributes. Overall, the study finds that LLMs demonstrate robustness across race and gender when it comes to matching resumes with job categories. However, variations in performance are observed for pregnancy status and political affiliation. To uncover potential sources of bias within open-source LLMs used in the study, contrastive input decoding is employed. This technique helps shed light on any underlying biases that may exist within these models. In conclusion, this research contributes to our understanding of algorithmic hiring bias in the context of large language models. It highlights both the strengths and limitations of LLMs when it comes to matching resumes with job categories while considering protected attributes such as gender, race, maternity status, pregnancy status, and political affiliation. The findings provide valuable insights for improving fairness and mitigating biases in algorithmic hiring processes.