The Impact of Disability Disclosure on Fairness and Bias in LLM-Driven Candidate Selection

Gene Louis Kim; Mahammed Kamruzzaman

REVIEW 3 major objections 6 minor 17 references

The Impact of Disability Disclosure on Fairness and Bias in LLM-Driven Candidate Selection

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper reports that, with qualifications held identical, LLMs preferred candidates who disclosed 'no disability,' and even candidates who declined to answer were selected less often than those who said 'no.'

desk verdict A clean, well-scoped measurement of disclosure-phrasing bias in LLM hiring recommendations; treat the abstract's generality with caution but send it to review. read the letter →

arxiv 2506.00256 v1 pith:LDUN4MS4 submitted 2025-05-30 cs.CL

Mahammed Kamruzzaman , Gene Louis Kim This is my paper · ORCID

classification cs.CL

keywords disabilitydisclosureLLMhiringbiascandidateselectionAIfairnessresumescreeningintersectionalprefernottoanswerchi-squaredtest

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tests whether the wording of voluntary disability disclosure changes which candidates large language models select for a job. In Experiment 1, the authors built triplets of candidates who shared the same name, gender, race, qualifications, education, and experience, and varied only the disclosure answer: 'Yes, I have a disability or have had one in the past,' 'No, I do not have a disability and have not had one in the past,' and 'I do not want to answer.' Across five LLMs, sixteen occupations, and over one hundred thousand prompts, models consistently chose the candidate who said 'no disability' more often than either other profile, with all differences significant by chi-square tests. The paper's point is that LLM hiring tools appear to penalize disability disclosure, and even non-disclosure, on the basis of phrasing alone rather than job-relevant merit.

What carries the argument

The machinery is a matched-triplet comparison: three candidate profiles that are word-for-word identical in name, gender, race, and stated qualifications, differing only in the disability-disclosure answer. The same selection prompt is repeated over 16 occupations with near-parity employment rates for disabled and non-disabled workers, randomized candidate order, three phrasing variants, and 102,000 samples, with chi-square tests quantifying whether selection counts deviate from an equal split. This design isolates the causal effect of disclosure wording from every other candidate attribute.

What would settle it

A falsifying observation would be a real or simulated resume-screening run, using full resumes plus a separate voluntary-disclosure form or candidates with realistic differences in experience, in which the 'no disability' preference shrinks to chance or reverses across the same five models; if that occurs, the list-format triplet result does not generalize to operational hiring pipelines.

Watch

Extended reading notes

Core claim

The central discovery is that LLM-driven candidate selection is systematically skewed by disability disclosure phrasing. When profiles were otherwise identical, all five tested models—GPT4o-mini, Gemma2-9B, Mistral-7B, Qwen2.5-7B, and Llama3.2-3B—selected the explicitly 'no disability' candidate at the highest rate, in every occupation studied, and the effect was statistically significant. Candidates who chose 'I do not want to answer' were also less likely to be selected than those who explicitly said 'no,' though the ordering relative to disclosed disability varied by model. Follow-up experiments varying gender or race, with disability status held constant, show that disclosure also shifts which gender and racial groups are favored, with models often justifying choices by appealing to diversity while actually favoring one group disproportionately.

Load-bearing premise

The load-bearing premise is that asking a model to 'select one candidate' from a short list of otherwise identical profiles captures how LLM hiring tools actually encounter disability disclosure, rather than being an artificial format that overstates the effect.

Editorial extensions

If this is right

If these results hold in operational settings, LLM-based screening tools penalize candidates who disclose a disability even when qualifications are identical to other candidates.
The 'prefer not to answer' option does not protect candidates; in the tested setup it also lowered selection rates relative to stating 'no disability.'
The bias appears across all five tested models and across all 16 occupations, not as a quirk of one model or one job type.
Gender and race effects shift when disability information is present: for example, Qwen showed no significant race difference in some occupations with disability information but significant differences without it.
To make LLM hiring fair, disability disclosure status should not be part of the candidate representation the model scores on.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A consequence that follows but is not tested here: applicants face a disclosure dilemma, because both 'yes' and 'I do not want to answer' carry a measured selection penalty relative to 'no.'
The chosen model justifications suggest the effect may stem from pretraining associations between disability and accommodation needs or lower adaptability; directly measuring those associations would be a natural follow-up.
The same matched-triplet protocol could be applied to other voluntarily disclosed protected attributes, such as veteran status; comparable results would show the bias is not disability-specific.
It remains open whether the bias can be removed by instruction prompts that tell the model to ignore disclosure information; that is a concrete, testable extension of the paper's setup.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Bottom line: this is a clean, well-scoped measurement showing that when candidates are otherwise identical, LLMs systematically prefer explicit 'no disability' over 'prefer not to answer' and over 'yes disability', across five models and sixteen occupations. The new bit is the 'prefer not to answer' penalty; prior work mostly looked at disability-related resume text. The study does what it claims, and the data and code are public.

What it does well: large samples (102k for Experiment 1), consistent chi-square results, an occupation set chosen to avoid a confounding employment gap, and a sensible baseline prompt with two rephrasings. The qualitative examples are few but they do show the mechanism: models reasoning that a candidate without disability is more adaptable, more confident, or less of an accommodation burden. That is real evidence of stereotype propagation, not just a statistical artifact.

The biggest soft spot is ecological validity. In Experiment 1, the candidate profiles contain no actual qualifications, education, or experience. The prompt tells the model the candidates are similar, but the model never sees any such content. So the result is a forced-choice tie-break under artificial parity, not a realistic resume screen. That is exactly what the paper sets out to measure, so I would not call it fatal, but the abstract's language about 'LLM-driven candidate selection' overreaches. A more precise claim would be 'when a model is asked to choose among candidates who differ only in disclosure phrasing.' Also, the paper gives chi-square and p-values but no effect sizes or confidence intervals; with 100k samples, almost anything is significant. A Cramer's V or similar would help. Minor issues: no sampling configuration or commit hash for exact replication, and the qualitative analysis is a handful of examples, fine as illustration.

Who this is for: fairness researchers, HR-tech auditors, and anyone building LLM evaluation suites. It is a useful, reproducible measurement protocol. I would not cite it as proof that real hiring pipelines discriminate this way, but I would cite it as evidence that models can behave this way under controlled conditions.

Recommendation: send it to peer review. Ask for effect sizes, a sentence or two on the forced-choice limitation, and a small robustness check with profiles that include at least one qualification field to see whether the preference survives when the model has actual job-relevant content to weigh.

Referee Report

3 major / 6 minor

Summary. The paper studies whether LLM-based candidate selection is affected by disability disclosure. In Experiment 1, three candidate profiles are identical except for the disability-disclosure response ("Yes, I have a disability", "No, I do not have a disability", and "I do not want to answer"), and five LLMs are asked to select one candidate for 16 occupations with small employment-rate gaps between persons with and without disabilities. The authors report large chi-square statistics and conclude that the models consistently favor the "no disability" candidate over both the "disability" and the "do not want to answer" candidates. Experiments 2 and 3 vary gender and race under a fixed "Yes" disability disclosure and compare with a condition where disability information is absent, reporting additional gender and race effects. Qualitative examples of model justifications are provided, and the code and data are publicly released.

Significance. If the central claim holds, the finding is practically important: identical candidates would be selected differently solely on the basis of the wording of a voluntary disclosure item, which has direct implications for AI-assisted hiring and for applicant guidance. Strengths include the large sample sizes, the use of five different models, the externally sourced name-race-gender dataset, the explicit control of candidate attributes, and the public release of artifacts. The paper also states its normative fairness baseline openly. However, the experimental format is highly stylized, and the statistical reporting does not currently support the strength of the claims. The main contribution is a controlled measurement of tie-breaking behavior rather than a demonstration of how real LLM-driven pipelines process disclosure in the presence of varied qualifications.

major comments (3)

[Experimental Setup, Experiment 1; Table 1] The candidate profiles in Table 1 contain only name, gender, race, and the disability-response line; the prompt's assertion that "candidates possess similar qualifications, education, and experience" gives the model no actual qualification content to process. The experiment therefore measures how an LLM breaks a tie when the only varying attribute is the disclosure statement, not whether disclosure would override qualifications in a realistic resume. This gap is load-bearing for the abstract's broad conclusion about "LLM-driven candidate selection." I request either (a) a condition with realistic resume content, including varied qualifications, or (b) a narrowed claim stating that the result concerns forced-choice tie-breaking under artificial parity. Additionally, there is no condition in which the disability field is entirely absent; only "I do not want to answer" is used, so the comparison to non-disclosure is incomplete.
[Results and Discussion, Table 2] All Experiment 1 significance claims are based on chi-square tests with sample sizes around 102k, so p < 0.05 is practically guaranteed and does not indicate the magnitude of the preference. The manuscript should report effect sizes (e.g., Cramér's V) or selection proportions with confidence intervals. The claim that the trend is "consistent across all occupations" is not checkable because no per-occupation counts or tests are reported for Experiment 1. If per-occupation tests were performed, a multiple-comparison correction is required. Also, p-values should be reported as "< 0.0001" rather than "0.0000".
[Experiment 2 and Experiment 3; RQ2/RQ3] The design compares a "Yes, I have a disability" profile with a condition in which disability information is entirely absent, but it does not cross the three disclosure responses with gender or race. As a result, RQ2 and RQ3, as stated in the Introduction, are not answered: the experiments show effects of adding versus omitting a disability statement, not the intersection of disclosure choice with gender/race. The text should either be reframed accordingly or expanded to include the full crossing.

minor comments (6)

[Table 1] Table 1 is difficult to parse as formatted; a table with separate rows for each profile and aligned fields would improve readability.
[Figures 1-3] Figures 1-3 omit axis labels and error bars; they should report the underlying proportions or counts, and the y-axis should be labeled.
[Desired Behavior of LLMs] The "Desired Behavior of LLMs" section presents neutrality as the expected behavior; this is a normative choice and should be labeled as such rather than as a property of the models.
[Experimental Setup] The paper says "We use all the default parameters" but does not list the specific sampling settings, temperature, or decoding parameters; reporting these would improve reproducibility.
[Throughout] There are several typos and spacing issues (e.g., "disability disclosure, influences" in the Introduction, and "University of South Florida" in the acknowledgments); the manuscript needs a careful proofreading pass.
[References] The reference list includes inconsistent author-name formatting (e.g., "Daum ´e Hii" in Nghiem et al.); please align with the venue style.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is a direct empirical measurement of LLM selection behavior with no fitted parameters, no self-citation chain, and no definitional reduction.

full rationale

The paper's central claim is an experimentally measured behavioral tendency: three otherwise identical candidate profiles differing only in disability-disclosure wording are presented to five LLMs, and the selection frequencies are tabulated. There is no equation that derives the outcome from the input, no fitted parameter later renamed as a prediction, and no load-bearing citation to the authors' own prior work; the reference list contains no self-citations. The 'desired behavior' baseline, uniform selection given identical qualifications, is a stated normative criterion rather than a fitted target. The choice of occupations with less than two percent employment-rate gaps and the use of an external name dataset are design controls, not post hoc fits. The skeptic's concern that the stripped-down candidate profiles lack substantive qualification content is an ecological-validity critique, not circularity: the experiment transparently reports the profile format and the prompt's assertion of identical qualifications. Accordingly, no circular step can be quoted or exhibited, and the score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No numerical free parameters are fitted to data. The study introduces no new theoretical entities. The central empirical claim depends on the validity of the name-based demographic signals, the normative fairness baseline, and the occupation selection criterion.

assumptions (3)

domain assumption The candidate names from Nghiem et al. (2024) carry the intended race and gender signals as perceived by the five LLMs.
Experiments 2 and 3 vary race or gender by name; if models do not perceive the intended identity, those results are unreliable.
domain assumption The fairness baseline is equal selection rates across disclosure statuses when qualifications are identical.
The paper defines the desired behavior of LLMs as neutrality toward disability disclosure; this normative choice underlies the interpretation of deviation as bias.
domain assumption The Bureau of Labor Statistics employment-rate-gap criterion (under 2%) selects occupations where disability status is not a meaningful occupational confound.
The paper uses this criterion to prevent job type from confounding disability effects; if the criterion is insufficient, the disability effect could be occupation-specific.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Impact of Disability Disclosure on Fairness and Bias in LLM-Driven Candidate Selection." pith.science (2026). https://pith.science/paper/LDUN4MS4

@misc{pith2026250600256,
  author       = {Pith},
  title        = {Pith review of: The Impact of Disability Disclosure on Fairness and Bias in LLM-Driven Candidate Selection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LDUN4MS4}},
  note         = {Machine review of arXiv:2506.00256}
}

read the original abstract

As large language models (LLMs) become increasingly integrated into hiring processes, concerns about fairness have gained prominence. When applying for jobs, companies often request/require demographic information, including gender, race, and disability or veteran status. This data is collected to support diversity and inclusion initiatives, but when provided to LLMs, especially disability-related information, it raises concerns about potential biases in candidate selection outcomes. Many studies have highlighted how disability can impact CV screening, yet little research has explored the specific effect of voluntarily disclosed information on LLM-driven candidate selection. This study seeks to bridge that gap. When candidates shared identical gender, race, qualifications, experience, and backgrounds, and sought jobs with minimal employment rate gaps between individuals with and without disabilities (e.g., Cashier, Software Developer), LLMs consistently favored candidates who disclosed that they had no disability. Even in cases where candidates chose not to disclose their disability status, the LLMs were less likely to select them compared to those who explicitly stated they did not have a disability.

Figures

Figures reproduced from arXiv: 2506.00256 by the authors.

**Figure 2.** Model-wise results for candidate selection of Experiment 2. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Candidate selection of LLMs for Experiment 3. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 15 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...
[2]

Beatty, D.; Masanthia, K.; Kaphol, T.; and Sethi, N. 2024. Revealing hidden bias in ai: Lessons from large language models. arXiv preprint arXiv:2410.16927

arXiv 2024
[3]

J.; Beltran, J

Budhwar, P.; Chowdhury, S.; Wood, G.; Aguinis, H.; Bamber, G. J.; Beltran, J. R.; Boselie, P.; Lee Cooke, F.; Decker, S.; DeNisi, A.; et al. 2023. Human resource management in the age of generative artificial intelligence: Perspectives and research directions on chatgpt. Human Resource Management Journal 33(3):606--659

work page 2023
[4]

Charmaz, K. 2010. Disclosing illness and disability in the workplace. Journal of International Education in Business 3(1/2):6--19

work page 2010
[5]

Evans, H. D. 2019. ‘trial by fire’: forms of impairment disclosure and implications for disability identity. Disability & Society 34(5):726--746

work page 2019
[6]

i wouldn’t say offensive but

Gadiraju, V.; Kane, S.; Dev, S.; Taylor, A.; Wang, D.; Denton, E.; and Brewer, R. 2023. " i wouldn’t say offensive but...": Disability-centered perspectives on large language models. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency , 205--216

work page 2023
[7]

Glazko, K.; Mohammed, Y.; Kosa, B.; Potluri, V.; and Mankoff, J. 2024. Identifying and improving disability bias in gpt-based resume screening. In The 2024 ACM Conference on Fairness, Accountability, and Transparency , 687--700

work page 2024
[8]

E., and Nikulin, M

Greenwood, P. E., and Nikulin, M. S. 1996. A guide to chi-squared testing , volume 280. John Wiley & Sons

work page 1996

Show all 17 references

[9]

Harris, C. 2023. Mitigating age biases in resume screening ai models. In The International FLAIRS Conference Proceedings , volume 36

2023
[10]

Kodiyan, A. A. 2019. An overview of ethical issues in using ai systems in hiring with a case study of amazon’s ai based hiring tool. Researchgate Preprint 1--19

2019
[11]

J.; Martinez, L

Lyons, B. J.; Martinez, L. R.; Ruggs, E. N.; Hebl, M. R.; Ryan, A. M.; O’Brien, K. R.; and Roebuck, A. 2018. To say or not to say: Different strategies of acknowledging a visible disability. Journal of Management 44(5):1980--2007

2018
[12]

what should i say to my employer… if anything?

Marshall, J. E.; Fearon, C.; Highwood, M.; and Warden, K. 2020. “what should i say to my employer… if anything?”-my disability disclosure dilemma. International Journal of Educational Management 34(7):1105--1117

2020
[13]

Nghiem, H.; Prindle, J.; Zhao, J.; and Daum \'e Iii, H. 2024. you gotta be a doctor, lin : An investigation of name-based bias of large language models in employment recommendations. In Al-Onaizan, Y.; Bansal, M.; and Chen, Y.-N., eds., Proceedings of the 2024 Conference on Em...

2024
[14]

W.-H.; Al-Emran, M.; Al-Sharafi, M

Ooi, K.-B.; Tan, G. W.-H.; Al-Emran, M.; Al-Sharafi, M. A.; Capatina, A.; Chakraborty, A.; Dwivedi, Y. K.; Huang, T.-L.; Kar, A. K.; Lee, V.-H.; et al. 2023. The potential of generative artificial intelligence across disciplines: Perspectives and future directions. Journal of ...

2023
[15]

Rane, N. 2023. Role and challenges of chatgpt and similar generative artificial intelligence in human resource management. Available at SSRN 4603230

2023
[16]

N.; Srinath, M.; and Wilson, S

Venkit, P. N.; Srinath, M.; and Wilson, S. 2022. A study of implicit bias in pretrained language models against people with disabilities. In Proceedings of the 29th International Conference on Computational Linguistics , 1324--1332

2022
[17]

Wilson, K., and Caliskan, A. 2024. Gender, race, and intersectional bias in resume screening via language model retrieval. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , volume 7, 1578--1590

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

[2] [2]

Beatty, D.; Masanthia, K.; Kaphol, T.; and Sethi, N. 2024. Revealing hidden bias in ai: Lessons from large language models. arXiv preprint arXiv:2410.16927

arXiv 2024

[3] [3]

J.; Beltran, J

Budhwar, P.; Chowdhury, S.; Wood, G.; Aguinis, H.; Bamber, G. J.; Beltran, J. R.; Boselie, P.; Lee Cooke, F.; Decker, S.; DeNisi, A.; et al. 2023. Human resource management in the age of generative artificial intelligence: Perspectives and research directions on chatgpt. Human Resource Management Journal 33(3):606--659

work page 2023

[4] [4]

Charmaz, K. 2010. Disclosing illness and disability in the workplace. Journal of International Education in Business 3(1/2):6--19

work page 2010

[5] [5]

Evans, H. D. 2019. ‘trial by fire’: forms of impairment disclosure and implications for disability identity. Disability & Society 34(5):726--746

work page 2019

[6] [6]

i wouldn’t say offensive but

Gadiraju, V.; Kane, S.; Dev, S.; Taylor, A.; Wang, D.; Denton, E.; and Brewer, R. 2023. " i wouldn’t say offensive but...": Disability-centered perspectives on large language models. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency , 205--216

work page 2023

[7] [7]

Glazko, K.; Mohammed, Y.; Kosa, B.; Potluri, V.; and Mankoff, J. 2024. Identifying and improving disability bias in gpt-based resume screening. In The 2024 ACM Conference on Fairness, Accountability, and Transparency , 687--700

work page 2024

[8] [8]

E., and Nikulin, M

Greenwood, P. E., and Nikulin, M. S. 1996. A guide to chi-squared testing , volume 280. John Wiley & Sons

work page 1996