Can AI Adoption Mitigate LGBTQ-Induced Discrimination in Hiring? Experimental Evidence from China

How to cite

WU, Xinyu; ZHANG, Xu. 2025. "Can AI Adoption Mitigate LGBTQ-Induced Discrimination in Hiring? Experimental Evidence from China." AEA RCT Registry. AEARCTR-0015759. https://doi.org/10.1257/rct.15759-2.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Xinyu WU* , Xu ZHANG

* Primary investigator

Registry ID
AEARCTR-0015759
DOI
https://doi.org/10.1257/rct.15759-2.0
First registered
Last updated
Status
On going
Trial dates
2025-04-09 – 2025-10-09
Countries
China ()
JEL codes
J71, O33, J16, D91
Registry record
http://www.socialscienceregistry.org/trials/15759

Abstract

LGBTQ+ individuals and their broader supporters often face significant barriers during the job application process. The growing adoption of Artificial Intelligence (AI) in recruitment presents a potential solution to this issue. Through two experiments, this study investigates whether AI adoption can mitigate LGBTQ-induced discrimination in hiring. The first experiment uses correspondence tests with fictitious resumes to quantify the impact of LGBTQ-related signals on hiring outcomes, uncovering biases faced by LGBTQ+ individuals and their allies. The second experiment evaluates AI’s effectiveness in reducing discrimination by comparing traditional, AI-assisted, and fully AI-driven recruitment models. Using a field experiment involving simulated HR professionals, we analyze how AI influences bias in decision-making. The findings reveal AI's dual-edged nature in recruitment, highlighting its ability to reduce human biases while raising concerns about potential algorithmic biases. This study contributes to understanding LGBTQ+-related discrimination and provides actionable insights for promoting fairness and inclusivity in AI-driven recruitment.

Intervention

In the first experiment, we designed matched triplets of fictitious resumes for various job positions. Each triplet contains three versions that are identical except for the LGBTQ-related signal: one resume includes an "LGBTQ signal," one includes a "Radical signal" (used as a mechanism), and one is a "Neutral" resume. These signals are conveyed through descriptions of organizational activities and work experience.

In the second experiment, we randomized HR evaluators into three groups: traditional, AI-assisted, and fully AI-driven recruitment. Each evaluator was given a set number of resumes (including resumes with the three types of signals) to score and evaluate.

Experimental design

(1) Experimental Design1:
To examine LGBTQ-related discrimination in the hiring process, we employ an audit study, a widely used method in discrimination research:
Step 1: Design a set of fictitious resumes (using accounting positions as an example), consisting of three resumes. The first resume includes an LGBTQ-signal, the second contains an Radical signal in the corresponding section, and the third is a Neutral Resume. Apart from the signal section, all three resumes are identical (e.g., age, educational background).
Step 2: Submit a large number of applications for accounting job postings via recruitment platforms, like Zhilian Zhaopin. However, ensure that no applications are submitted to the same company.
Step 3: Test using the regression model. [Callbacki = α + β1LGBTQi + ϵi]
In this model, Callbacki is a binary variable indicating whether resume i received a callback for an interview, while LGBTQi is a binary variable indicating whether resume i contains LGBTQ-related signals. The coefficient β1 measures the extent to which LGBTQ-related signals influence the probability of receiving a callback. If β1 is significantly negative, it indicates the presence of discrimination against LGBTQ-related signals in the hiring process.

(2) Experimental Design2:
Step1: Recruitment of Evaluators: We plan to recruit 500 professionals from accounting and finance fields with hiring experience. Each evaluator will be paid ¥50 to assess resumes, and they will be informed that their evaluations directly affect candidates’ opportunities.
Step2: Group Division: Evaluators will be divided into three groups: (1)Traditional Recruitment Group (2) AI-Assisted Group: Evaluates resumes with AI-generated compatibility scores as reference. (3) Pure AI Decision Group.
Step 3: Collect Resume: A total of 1,250 resumes will be collected. Among these, 125 resumes (10%) will be modified to include an LGBTQ-signal, resulting in a dataset with 125 LGBTQ-signal resumes and 1,125 regular resumes.
Step 4: Resumes will be randomly distributed to evaluators, with each evaluator reviewing five resumes (one of which contains an LGBTQ-signal).
Step 5: We will collect the scoring data and conduct a post-survey to examine the mechanisms behind the evaluators’ decisions.
Step 6: We will analyze the results using regression analysis, controlling for several factors, including applicant characteristics such as gender, age, GPA, total months of relevant internship experience, computer skills, volunteering experience, marital status, and whether the applicant has children. Additionally, evaluator fixed effects will be included to account for unobserved heterogeneity among evaluators.

Design details

Randomization method
(1) In the fictitious resume experiment, each resume is randomly assigned to a specific job description. (2) In the second field experiment, HR evaluators are randomly grouped, and the resumes to be scored are randomly assigned.
Randomization unit
"Individual resumes (Experiment 1); Individual HR evaluators (Experiment 2)."
Clusters
2100 fictitious resumes (Experiment 1); 500 HR evaluators and 1250 resumes (Experiment 2)
Observations
2000 fictitious resumes (Experiment 1); 500 HR evaluators and 1250 resumes (Experiment 2)
Arms
Experiment 1: 700 resumes - Treatment group (LGBTQ), 700 resumes - Control group1 (radical), 700 resumes - Control group2 (neutral) Experiment 2: 250 HR Treatment group (AI-assist score), 250 HR Control group1 (Traditional score), Purely-AI score do not need HR; Each HR evaluator gives a score for 5 different resumes
Primary outcomes
(1) Experiment 1: Callback rate for resumes with "LGBTQ signals" vs. "Radical resumes" vs. "Neutral resumes." (2) Experiment 2: HR evaluation scores (0-100) under Traditional evaluation vs. AI-assisted vs. Pure AI.

Cited by

No citation data available for this registration yet.

Keywords