Large language models (LLMs) are increasingly integrated into AI-assisted hiring pipelines, including automated resume generation and screening. Under the EU AI Act, the hiring domain is classified as high-risk, making fairness and transparency critical requirements. Existing work has primarily focused on explicit hiring decisions, while less attention has been paid to whether generated resumes themselves encode recoverable demographic information. In this work, we conduct a two-stage audit of demographic leakage in German-language LLM-generated resumes. First, we use ChatGPT (GPT-4o-mini), Gemini 2.5 Flash-Lite, and multiple scales of the open-weight Qwen 3 model family (4B, 8B, and 14B) to generate resumes from real anonymized job-matching profiles, systematically varying gender- and ethnicity-associated names while holding qualifications constant. Second, we simulate a downstream resume screening scenario, where the generated resumes are first anonymized and gender-neutralized, before demographic leakage classifiers are trained on the resulting texts. We find that, despite these interventions, classifiers reliably distinguish between resumes generated with male and female names. This leakage is not driven by overtly gendered wording, but by subtle differences in the usage of semantically equivalent, formally gender-neutral terms in German. In contrast, ethnicity-related leakage remains comparatively weak across models. Our findings demonstrate that apparently neutral resume generation can still preserve highly predictive demographic signals, raising concerns about anonymization-based fairness interventions in multilingual AI hiring pipelines.
Figures & tables
Figure 1: Overview of the two-stage audit pipeline. Stage 1 generates résumés from real anonymized job-matching profiles by systematically varying names as proxies for gender and ethnicity. Stage 2 assesses demographic leakage by anonymizing and gender-neutralizing the résumés before evaluating gender and ethnicity recoverability using TF-IDF representations and logistic regression classifiers.
Original word
Gender-neutral form
Beraterin ( female consultant )
→
Berater ( consultant )
Kaufmann ( salesman )
→
Kauf-mann/-frau ( salesperson )
Kauffrau ( saleswoman )
→
Kauf-mann/-frau ( salesperson )
Fachmann ( male specialist )
→
Fach-mann/-frau ( specialist )
Fachfrau ( female specialist )
→
Fach-mann/-frau ( specialist )
Table 1: Examples of gender-stem normalization.
Figure 2: Distribution of total résumé length (word count) across LLM families (left) and Qwen 3 model sizes (right).
Term
n (/30)
Mean
Mean selected
Performance
coefficient
features
Acc
AUC (sd)
ChatGPT
4.9
0.825
0.900 (0.046)
fachkraft specialist n.
30
-0.991 (f.)
fach-mann/-frau specialist m./f.
30
0.600 (m.)
Gemini
6.9
0.877
0.915 (0.034)
fachkraft specialist n.
30
-1.541 (f.)
Table 2: Most frequently selected TF–IDF lasso terms for gender prediction across 30 repeated grouped LASSO runs per model. Per term, we report the corresponding number of selected runs and mean coefficient across runs. Per model, we report mean selected features per run and mean classifier performance (Accuracy and AUC, with standard deviations (sd) over all runs).
Term
n (/30)
Mean
Mean selected
Performance
coefficient
features
Acc
AUC (sd)
ChatGPT
1.7
0.500
0.499 (0.005)
muttersprache mother tongue
16
-0.001 (t.)
Gemini
3.1
0.521
0.540 (0.027)
austausch exchange
24
-0.023 (t.)
Qwen 8B top-k = 20
1.6
0.500
0.503 (0.012)
Table 3: Most frequently selected TF–IDF lasso terms for ethnicity prediction across 30 repeated grouped LASSO runs per model. Per term, we report the corresponding number of selected runs and mean coefficient across runs. Per model, we report mean selected features per run and mean classifier performance (Accuracy and AUC, with standard deviations (sd) over all runs).
Target
Setting
4B
8B
14B
Acc
AUC
Acc
AUC
Acc
AUC
gender
top-k=20
0.636
0.683
0.670
0.754
0.687
0.686
top-k=64
0.616
0.652
0.654
0.704
0.670
0.679
ethnicity
top-k=20
0.500
0.499
0.500
0.503
0.566
0.579
top-k=64
0.519
0.529
0.500
0.500
0.542
0.557
Table 4: Leakage classifier performance across Qwen 3 model sizes and top-k decoding settings. Performance is averaged across 30 repeated runs.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Item Name
Type
Answers
Subanswers
Hard Skills
Industry Experience
MULTI
Open to all; Agency / Advertising / Marketing / PR; Plant and mechanical engineering / industry; …
Table 5: Overview of matching profile questionnaire items. Translated from German; illustrative selection of response options. Items are grouped into hard skills, soft skills, and preferences.
German
Turkish
Male
Female
Male
Female
Marcel Popp
Ramona Popp
Emre Eren
Havva Eren
Philipp Stoll
Mareen Stoll
Onur Ayyildiz
Kübra Ayyildiz
Kai Lindner
Carina Lindner
Orhan Turan
Cansu Turan
Fabian Kern
Britta Kern
Fikret Akman
Necla Akman
Marco Neubauer
Jennifer Neubauer
Yunus Sönmez
Gül Sönmez
Appendix
Table 6: Complete list of sampled names grouped by ethnicity and gender.
Figure 3: Distribution of total résumé length (word count) by gender across all model families. No systematic differences in résumé length are observed between male- and female-associated résumés.
Figure 4: Distribution of total résumé length (word count) by ethnicity across all model families. Résumé length remains broadly comparable between German- and Turkish-associated résumés, indicating no substantial ethnicity-related differences in output length.
Figure 5: Distribution of agentic word counts by gender across all model families.
Figure 6: Distribution of communal word counts by gender across all model families.
Algorithmic fairness studies how automated decisions distribute errors, opportunities, and harms across groups, and how to measure and reduce disparities.
De-identified résumé screening assumes that redacting explicit fields prevents ethnocultural inference; however, recent audits attribute residual leakage to declared languages. We investigate whether eliminating language fields resolves this leakage across nine open-weight models and 620 counterfactual résumés. By holding language attributes strictly identical, we isolate unstructured prose across five ethnocultural conditions and three cue-salience tiers. Target-group recovery averages 0.757 overall and saturates at 1.000 under high salience, demonstrating that non-language prose sustains demographic inference. Crucially, models diverge only under faint cues (0.086-0.690), establishing salience as an essential evaluation axis. Furthermore, pairwise LLM-as-a-judge outcomes are highly sensitive to evaluation design: forbidding ties yields an apparent selection-rate ratio of 0.39 alongside strong position and content effects, whereas permitting ties produces near-universal ties for most models (≥94%). Downstream scoring shows only very small between-condition differences, highlighting the need to distinguish demographic signals recoverable from résumé content from effects introduced by the evaluation protocol.
Qiangju Chen, Yang Xiao
Macquarie University · The University of Melbourne
We audit fourteen mainstream large language models (LLMs) for hiring discrimination using the paired-resume methodology of Kline, Rose, and Walters (2022). The sole 2023-vintage model reproduces the pro-White callback gap documented in field experiments on labor market discrimination (+2.12 pp, significant at the 1% level). Every model released in 2024 or after shows either a null gap or a significant pro-Black reversal (up to −3.01 pp). The same pattern holds on the gender axis. Based on 24,024 paired postings per model across 14 models, our results document a reversal in the direction of algorithmic hiring bias across model generations.
Zhenyu Gao, Wenxi Jiang, Yutong Yan
Department of Finance, CUHK Business School, The Chinese University of Hong Kong
We present PopResume, a population-representative resume dataset for causal fairness auditing of LLM- and VLM-based resume screening systems. Unlike existing benchmarks that rely on manually injected demographic information and outcome-level disparities, PopResume is grounded in population statistics and preserves natural attribute relationships, enabling path-specific effect (PSE)-based fairness evaluation. We decompose the effect of a protected attribute on resume scores into two paths: the business necessity path, mediated by job-relevant qualifications, and the redlining path, mediated by demographic proxies. This distinction allows auditors to separate legally permissible from impermissible sources of disparity. Evaluating four LLMs and four VLMs on PopResume's 60.8K resumes across five occupations, we identify five representative discrimination patterns that aggregate metrics fail to capture. Our results demonstrate that PSE-based evaluation reveals fairness issues masked by outcome-level measures, underscoring the need for causally-grounded auditing frameworks in AI-assisted hiring.
Sumin Yu, Juhyeon Park, Taesup Moon
ECE, Seoul National University · IPAI, Seoul National University · ASRI / INMC / AIIS, Seoul National University