Pith. sign in

REVIEW 3 major objections 4 minor 47 references

Current chat AI models carry encoded ableist bias toward people with intellectual disabilities, measurable in the stories they generate.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 10:58 UTC pith:IJARE72Q

load-bearing objection A transparent and plausible map of ableist tropes in LLM-generated stories, but the unvalidated, non-blinded GPT-4-Turbo judge makes the headline odds ratios uninterpretable as measurements of the generators' bias. the 3 major comments →

arxiv 2607.26062 v1 pith:IJARE72Q submitted 2026-06-16 cs.CY cs.AIcs.CL

Identifying Implicit Bias in LLM-based Chat AI Toward People with Intellectual Disabilities

Karly V. Coffey , Gloria L. Krahn , John P. Hanley , Jacob E. Neely This is my paper
classification cs.CY cs.AIcs.CL
keywords implicit biasintellectual disabilityableismlarge language modelsstory generationLLM-as-judgerepresentational biasAI fairness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that large language models used in chat systems systematically portray people with intellectual disabilities (ID) as younger, more dependent, more in need of help, more paternalized, and more inspirational than people without that descriptor—even when the story prompt is otherwise identical. The authors argue these differences exceed what actual characteristics of ID would explain and instead reflect implicit ableist stereotypes embedded in the models' training data. This matters because LLM-generated content increasingly shapes workplaces, healthcare materials, and social narratives, and unexamined bias could perpetuate historical discrimination against people with ID. The paper offers both evidence and a scalable method: use one LLM to generate stories and another to judge them on predefined bias themes.

Core claim

On the paper's own terms, the central claim is that LLM-generated stories about people described as having intellectual disabilities differ from control stories in consistent, measurable ways. Using GPT-4-Turbo as a judge, the authors report strikingly elevated odds ratios for infantilization (OR ≈ 407), paternalistic tone (OR ≈ 137), perceived or treated as younger (OR ≈ 83), dependency (OR ≈ 42), and need for help (OR ≈ 17), alongside sharply reduced autonomy and decision-making. All five tested LLMs agree directionally. The authors interpret this pattern as encoded implicit bias—stereotyping that goes beyond clinically relevant features of ID and matches ableist themes documented in prior

What carries the argument

The method pairs story generation with story comprehension: ten neutral prompt stems (e.g., 'Tell me a short story about a person …') are run with and without an 'intellectual disabilities' descriptor across five chat models, producing 25,000 stories. A separate GPT-4-Turbo instance answers binary thematic questions about each story—covering paternalism, infantilization, autonomy, dependency, inspirational framing, and help-seeking—and odds ratios between the two descriptor groups quantify the representational gap. The judge's yes/no outputs are the quantitative carrier of the argument.

Load-bearing premise

The entire quantitative result rests on trusting a separate chat model's yes/no answers as valid measures of concepts like paternalism and infantilization; if that judge is itself biased by seeing the ID descriptor, the odds ratios may reflect the judge's stereotypes rather than the generators'.

What would settle it

Have human raters blind to the descriptor code a random sample of the ID-present and ID-absent stories for the same themes. If human coders find no systematic difference between the groups (or much smaller differences), the paper's conclusion would collapse. Alternatively, a probe: present the judge with identical stories where 'intellectual disabilities' is replaced by an irrelevant descriptor (e.g., 'wearing glasses'); if the judge still flags high paternalism, its responses are not valid measures.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, chat AI models cannot be treated as neutral about disability; they will tend to cast people with ID as perpetual children needing rescue, even in everyday scenarios like going to a restaurant.
  • Auditors and developers of LLM-based systems should include ID-specific bias tests; the effect sizes here (odds ratios in the hundreds) dwarf typical fairness metrics.
  • Users of LLM-generated text—in workplaces, schools, healthcare—may unknowingly propagate ableist tropes, reinforcing social exclusion and self-fulfilling expectations.
  • The generate-and-judge pipeline offers a scalable template for surfacing implicit bias in under-studied groups, though its validity depends on judge calibration.
  • The directional agreement across five different models suggests the bias is systemic to current internet-scale training rather than a quirk of one vendor.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct extension the authors leave implicit: the same pipeline could be applied to image generation, predicting that text-to-image models would also depict people with ID as younger and more dependent—an easily testable corollary.
  • The 'inspirational' and 'symbol' findings imply that sentiment-focused guardrails (which only filter negative content) would miss this bias, because the harm operates through objectifying positivity, not overt negativity.
  • If the judge's measurements are trusted, the age skew toward 20–30 in ID stories suggests training corpora over-represent people with ID as school-age or young adults—an empirical hypothesis that could be checked with corpus analysis.
  • The biggest open question is judge validity; a human-coding follow-up on a subsample would either corroborate the magnitudes or force a rescaling, but even a direction-only confirmation would sustain the core conclusion.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper investigates implicit bias toward people with intellectual disabilities (ID) in LLM-generated stories. The authors generate 25,000 short stories from five LLMs (GPT-4-Turbo, GPT-4o, Llama-3.3-70B, Claude-3.5-Sonnet, Mistral-Large) using ten neutral prompt stems, each with and without an ID descriptor. A separate GPT-4-Turbo instance answers a battery of yes/no thematic questions about each story (paternalism, infantilization, dependency, inspiration, burden, etc.). The paper reports large odds ratios for ID-present versus ID-absent stories, e.g., infantilization OR = 407.43 [234.69, 707.32], paternalism OR = 136.59, dependency OR = 42.32, and interprets these as evidence of encoded implicit bias in LLM-generated products. Detailed results are presented for GPT-4-Turbo, with a heatmap suggesting directional agreement across all five generators.

Significance. If the measurement were valid, the finding would be important: it would provide large-scale, reproducible evidence of ableist representational bias in mainstream chat models, extending prior work on disability bias in LLMs and offering a scalable audit methodology. The paper has genuine strengths: the data-generation design isolates the ID descriptor by using matched prompt stems, the code and data are publicly available, and the themes align with established ableism literature. However, the central quantitative claim is conditional on the validity of an unvalidated, non-blinded LLM judge. Appendix 3 contains direct examples of the judge labeling a positive, uplifting ID story as 'Burden: Yes,' 'Negative Perception: Yes,' and 'Cognitive Ability: Yes' based on the presence of the descriptor. The reported odds ratios may therefore reflect the judge's own stereotyping rather than measurable differences in the generated stories. The significance is real but currently unestablished.

major comments (3)
  1. [Methods, Story Comprehension; Results, Figs. 2–4] The outcome variable is produced by a separate GPT-4-Turbo instance that reads the full story, including the ID descriptor. The judge is itself an LLM trained on text containing ableist stereotypes and is from the same model family as one of the generators. The paper reports no human validation, inter-rater reliability, or calibration of the judge's binary labels. Appendix 3 provides direct evidence of unreliability: for the positive story about Charlie competing in a race, the judge returns 'Burden: Yes,' 'Negative Perception: Yes,' 'Cognitive Ability: Yes,' and 'Infantilization: Yes,' with explanations such as 'The text explicitly mentions Charlie has intellectual disabilities.' This shows the judge is responding to the descriptor, not objectively measuring story content. The central claim is about bias in generated products, not about judge bias; the current design cannot separate the
  2. [Results, Figure 3; Story Comprehension] The age finding—that ID-present stories depict younger characters—is partly manufactured by the judge. The paper states, 'when age was not explicitly stated, GPT-4-Turbo implied age based on the actions and way the individual was treated within the story.' Thus, the age approximation is an LLM inference, not an observed property of the generated text. The reported OR of 6.65 for ages below 30 versus above 30 may reflect the judge's age stereotypes for people with ID rather than the generators' output. The authors should separate explicit age mentions from inferred ages, or validate the judge's age inference against human annotators.
  3. [Results, Figure 5; Discussion] The claim that 'all five LLM models agree directionally' is supported only by a heatmap with no per-model odds ratios, confidence intervals, or sample sizes in the main text. Because the same unvalidated judge is used for all generators, cross-model agreement cannot disentangle generator bias from judge bias. Even if the judge were validated, the detailed analysis is only presented for GPT-4-Turbo; the supplementary material should include full per-model tables and, ideally, separate judge validation or sensitivity analyses for each generator family.
minor comments (4)
  1. [Appendix 1 / Table 1] The comprehension prompt includes many variables not reported in Table 1 or the Results, such as 'Intelligence,' 'Skilled Labor,' 'Collar Labor,' 'Burden,' and 'Ability Bias.' Clarify which variables were analyzed and whether the analysis was pre-specified, and report or justify omission of the others.
  2. [Appendix 3] The example JSON outputs include malformed entries, e.g., a missing value after 'Infantilization': and 'Why Infantilization': null, as well as several null fields. This suggests parsing or cleaning issues that should be documented in the methods or reproducibility materials.
  3. [Methods, Data Generation] The prompt stems include 'at a bar,' which may introduce age-related confounds for the ID group if the judge associates ID with younger ages. It would be helpful to discuss whether this situation could affect the age comparisons.
  4. [Limitations] The limitations paragraph acknowledges that 'further humans-in-the-loop analyses could add to the robustness' and that the work 'would have been stronger' with an inclusive research approach. Given the load-bearing role of the LLM judge, human validation is not a secondary enhancement but a necessary check on the central claim; the manuscript should either provide it or substantially temper the conclusions.

Circularity Check

2 steps flagged

LLM judge is not blinded to the ID descriptor; Appendix 3 shows labels driven by the phrase 'intellectual disabilities,' so reported ORs may measure the judge, not the generated stories.

specific steps
  1. other [Methods, Story Comprehension for Bias Detection and Analysis; Appendix 3, first ID story (Charlie)]
    "After running the story generation model, we utilized a separate instance of the GPT-4-Turbo LLM to answer questions about the stories from themes identified in the literature [8, 21, 22, 23, 24]. ... "Cognitive Ability": "Yes", "Why Cognitive": "The text explicitly mentions Charlie has intellectual disabilities." ... "Burden": "Yes", "Why Burden": "The story suggests his participation and needs require substantial support and adjustments from others.""

    The outcome variables — every yes/no answer that feeds the odds ratios — are produced by an LLM judge whose input is the full story containing the ID descriptor. The judge is not blinded to the condition being measured. Appendix 3 shows the judge answering 'Yes' to Cognitive Ability because the text 'explicitly mentions' intellectual disabilities, and answering 'Burden: Yes' for a positive story about Charlie by inferring burden from support needs. Thus the group difference in judge labels may be caused by the judge's reaction to the phrase 'intellectual disabilities' rather than by measurable differences in story content. Because the judge is itself an LLM of the same kind being tested, the analysis cannot separate bias in the generation products from bias in the measuring instrument. The

  2. other [Results, paragraph beginning 'All five LLM models agree directionally...']
    "All five LLM models agree directionally with results reported in Figure 2, Figure 3, and Figure 4 with only small deviation in odds ratios (Figure 5, Appendix 2)."

    This cross-model agreement is presented as evidence that the bias resides in the stories generated by all five LLMs. However, every story from all five generators was scored by the same single GPT-4-Turbo judge. If that judge is systematically responding to the presence of the ID descriptor, all five model sets will appear to agree directionally regardless of the actual content of their stories. The shared judge makes the agreement a consequence of the measurement artifact, not independent confirmation from five distinct generation systems. This supporting claim therefore reduces to the same non-blinded judge behavior rather than providing separate evidence of encoded bias.

full rationale

The derivation is an empirical measurement, not a formal derivation, so the self-citation and imported-uniqueness patterns do not arise. The load-bearing problem is the measurement instrument. Every quantitative result is the output of a separate GPT-4-Turbo judge reading the full generated story, which contains the ID descriptor whenever the condition is ID Present. The judge is not blinded, and the paper does not validate it for these disability themes against human ratings or against ID-redacted versions. Appendix 3 provides direct evidence that the judge uses the phrase 'intellectual disabilities' as a sufficient reason to answer 'Yes' to Cognitive Ability and infers 'Burden' even in a positive story. Hence the group differences feeding all odds ratios may be produced by the judge's own stereotyping rather than by the story text. The paper's own limitation states that an LLM could perpetuate ableist bias in analysis, and it chose the LLM anyway. Cross-model agreement is also produced by this single shared judge, so it does not independently corroborate the result. Because the central claim — that generation products encode bias — is not separable from judge bias in this design, the result is partially circular. External citations that LLMs can do qualitative coding [35-41] are for other tasks and do not establish validity for these specific ID-bias labels, especially given the Appendix 3 counter-example. Score 6 reflects one central, partial circularity rather than a fully self-referential derivation.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No invented entities. The free parameters are the judge model and analysis thresholds; the central statistical estimates are directly determined by the unvalidated judge's answers. The axioms are the implicit measurement validity assumptions. This is an empirical audit, not a derivation.

free parameters (3)
  • LLM judge model = GPT-4-Turbo (OpenAI, Dec 2024)
    The entire outcome dataset consists of this model's answers; no human validation establishes its accuracy. The choice of judge determines all reported odds ratios.
  • Age threshold for 'younger' comparison = 30 years
    The paper reports OR=6.65 for below-30 vs above-30; the threshold appears to be chosen post hoc based on the observed age distributions, and the judge's age estimates are partly inferred from behavior (Methods, Figure 3).
  • Stories per condition = 250
    Arbitrary sample size per condition; no power analysis or multiple-seed variance is reported.
axioms (5)
  • domain assumption The GPT-4-Turbo judge's binary answers are valid, reliable measures of the constructs (paternalism, infantilization, etc.).
    The paper uses a separate GPT-4-Turbo instance to answer 30 Yes/No questions per story, without any human-annotated calibration or inter-rater reliability check. Appendix 3 shows the judge labeling a positive story 'Burden: Yes' and 'Negative Perception: Yes', so this assumption is doubtful. Location: Methods – Story Comprehension.
  • domain assumption ID Absent stories function as a control group for people without ID.
    Acknowledged in Limitations: 'We presume that ID absent stories imply that the subject of the story does not have ID; however, differences in perceived ability support our interpretation.' This is an untested assumption; 'a person' could be any person.
  • domain assumption The 250 seeded generations per condition are independent, representative samples from each model.
    Temperature 1.0 and API calls are not fully deterministic; the seed only partially controls stochasticity. The paper does not report multiple seeds or variance across seeds.
  • domain assumption The thematic questions and categories derived from prior disability literature are appropriate for ID and LLM-generated text.
    Questions like 'Is this person a burden to family or society?' are value-laden and applied to short stories; no evidence that these categories transfer to LLM narrative text. Location: Methods, Table 1.
  • standard math Clopper-Pearson confidence intervals and odds ratios are appropriate statistical tools for this design.
    These are standard methods, but they do not correct for multiple comparisons across ~20 questions, and the assumption of independent observations is imperfect if generations share model-internal context.

pith-pipeline@v1.3.0-alltime-deepseek · 21972 in / 16685 out tokens · 151136 ms · 2026-08-02T10:58:20.864921+00:00 · methodology

0 comments
read the original abstract

Background: This work investigates the presence of implicit bias in Large Language Model (LLM)-based chat AI models directed toward people with intellectual disabilities (ID). Objective: The study aims to identify and measure representational differences related to people with ID and examine them to identify implicit biases inherent in AI chat generation technologies. Methods: Utilizing the GPT-4-Turbo model, we requested story-generation based on 10 prompt stems with and without descriptors for ID. This process was repeated using four other LLMs (OpenAI GPT-4o, Meta Llama-3-3-70B-Instruct, Anthropic Claude-3-5-Sonnet, and Mistral-Large-2411). The resulting 25,000 computer-generated stories were analyzed using a separate GPT-4-Turbo model instance to detect differences in how people are represented related to themes of bias described in previous literature. Results: Our findings reveal differences in how people are represented between story datasets with and without ID descriptors. These differences go beyond established characteristics of ID and imply the presence of mostly negative implicit biases. Identified differences related to considering people with ID as younger, with themes of paternalism and infantilization; depicting them as more inspirational and symbolic; as needing help more often, being dependent, and being saved; and having a negative perception of them and more hesitation to include them. Conclusions: These implicit biases are considered within the context of past discrimination towards people with ID and highlight the need for diligence against implicit bias towards people with ID in AI development. This research underscores the importance of assessing and mitigating implicit bias in decision-making technologies to prevent future societal harm.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 12 canonical work pages · 4 internal anchors

  1. [1]

    Shen, Y., Song, K., Tan, X., Zhang, W., Ren, K., Yuan, S., Lu, W., Li, D., & Zhuang, Y. (2024). TaskBench: Benchmarking large language models for task automation. arXiv. https://doi.org/10.48550/arXiv.2311.18760

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.18760 2024

  2. [2]

    Makhmutova, L., Salton, G., Perez-Tellez, F., & Ross, R. (2024). Automated medical text simplification for enhanced patient access. In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (pp. 208 –218). SCITEPRESS . https://doi.org/10.5220/0012466100003657

    work page doi:10.5220/0012466100003657 2024

  3. [3]

    (2025, February 25)

    Lin, L., & Parker, K. (2025, February 25). U.S. workers are more worried than hopeful about future AI use in the workplace. Pew Research Center . https://www.pewresearch.org/social-trends/2025/02/25/u-s-workers-are-more-worried-than-hopeful-about-future-ai-use-in-the-workplace/

    2025

  4. [4]

    (2023, June 8)

    Weiser, B., & Schweber, N. (2023, June 8). The ChatGPT lawyer explains himself. The New York Times . https://www.nytimes.com/2023/06/08/nyregion/lawyer-chatgpt-sanctions.html

    2023

  5. [5]

    Friedman, C. (2023). Explicit and implicit: Ableism of disability professionals. Disability and Health Journal. https://doi.org/10.1016/j.dhjo.2023.101482

    arXiv 2023

  6. [6]

    R., & Dunn, D

    Bogart, K. R., & Dunn, D. S. (2019). Ableism special issue introduction. Journal of Social Issues, 75(3), 650–664. https://doi.org/10.1111/josi.12354

    work page doi:10.1111/josi.12354 2019

  7. [7]

    I., Rao, S

    Iezzoni, L. I., Rao, S. R., Rsalam, J., Bolcic-Jankovic, D., Agaronnik, N. D., Donelan, K., Lagu, T., & Campbell, E. G. (2021). Physicians’ perceptions of people with disability and their health care. Health Affairs. https://doi.org/10.1377/hlthaff.2020.01452

    arXiv 2021

  8. [8]

    A., van Oorsouw, W

    Pelleboer-Gunnink, H. A., van Oorsouw, W. M. W. J., van Weeghel, J., & Embregts, P. J. C. M. (2019). Stigma research in the field of in tellectual disabilities: A scoping review on the perspective of care providers. International Journal of Developmental Di sabilities, 67(3), 168 –187. https://doi.org/10.1080/20473869.2019.1616990

    arXiv 2019

  9. [9]

    Friedman, C., & VanPuymbrouck, L. (2019). The relationship between disability prejudice and Medicaid home and community -based services spending. https://doi.org/10.1016/j.dhjo.2019.01.012

    work page doi:10.1016/j.dhjo.2019.01.012 2019

  10. [10]

    L., & Disability Rights Education and Defense Fund (DREDF)

    Yee, S., Breslin, M. L., & Disability Rights Education and Defense Fund (DREDF). (2023). This data, not that data: Big data, privacy, and the impact on people with disabilities. Retrieved January 8, 2025, from https://healthlaw.org/wp-content/uploads/2023/03/This-Data-Not-That-Data_Disability-Rights- Education-and-Defense-Fund_FINAL.pdf

    2023

  11. [11]

    Pareek, C. S. (2022). Unmasking bias: A framework for testing and mitigating AI bias in insurance underwriting models. Journal of AI, ML, and Data, 1(1), 1736–1741. https://doi.org/10.51219/JAIMLD/Chandra-shekhar-pareek/377

    work page doi:10.51219/jaimld/chandra-shekhar-pareek/377 2022

  12. [12]

    D., Gonzalez-Guarda, R., Lawrence, K., Patil, C

    Boyd, A. D., Gonzalez-Guarda, R., Lawrence, K., Patil, C. L., Ezenwa, M. O., O’Brien, E. C., Paek, H., Braciszewski, J. M., Adeyemi, O., Cuthel, A. M., Darby, J. E., Zigler, C. K., Ho, P. M., Faurot, K. R., Staman, K., Leigh, J. W., Dailey, D. L., Cheville , A., Del Fiol, G., Knisely, M. R., Marsolo, K., Richesson, R. L., & Schlaeger, J. M. (2023). Equity...

    arXiv 2023

  13. [13]

    Glazko, K., Mohammed, Y., Kosa, B., Potluri, V., & Mankoff, J. (2024). Identifying and improving disability bias in GPT -based resume screening. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24) (pp. 687 –700). ACM. https://doi.org/10.1145/3630106.3658933

    arXiv 2024

  14. [14]

    Manzoor, R., Hussain, W., & Anjum, M. L. (2024). Out of dataset, out of algorithm, out of mind: A critical evaluation of AI b ias against disabled people. AI & Society. https://doi.org/10.1007/s00146-024-02168-8

    work page doi:10.1007/s00146-024-02168-8 2024

  15. [15]

    O., Rossi, R

    Gallegos, I. O., Rossi, R. A., Barrow, J., Tanjim, M. M., Kim, S., Dernoncourt, F., Yu, T., Zhang, R., & Ahmed, N. K. (2024). Bias and fairness in large language models: A survey. Computational Linguistics, 50(3), 1097–1179. https://doi.org/10.1162/coli_a_00524

    doi:10.1162/coli_a_00524 2024

  16. [16]

    Dodge, J., Sap, M., Marasović, A., Agnew, W., Ilharco, G., Groeneveld, D., Mitchell, M., & Gardner, M. (2021). Documenting la rge webtext corpora: A case study on the colossal clean crawled corpus. arXiv. https://doi.org/10.48550/arXiv.2104.08758

    Pith/arXiv arXiv doi:10.48550/arxiv.2104.08758 2021

  17. [17]

    Rekabsaz, N., West, R., Henderson, J., & Hanbury, A. (2021). Measuring societal biases from text corpora with smoothed first-order co-occurrence. arXiv. https://doi.org/10.48550/arXiv.1812.10424 15

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1812.10424 2021

  18. [18]

    Navigli, R., Conia, S., & Ross, B. (2023). Biases in large language models: Origins, inventory, and discussion. Journal of Data and Information Quality, 15(2), 1–21. https://doi.org/10.1145/3597307

    doi:10.1145/3597307 2023

  19. [19]

    Muñoz-García, V. (2024). Bias mitigation in corpora for LLMs training applied to text simplification. Doctoral Symposium on Natural Language Processing (September 2024). Retrieved from https://ceur-ws.org/Vol-3797/paper5.pdf

    2024

  20. [20]

    (2025, March 5)

    Luse, B. (2025, March 5). The R-word is back. But why? NPR. https://www.npr.org/2025/03/05/1263527053/the-r-word-disability-ableism

    2025

  21. [21]

    I wouldn’t say offensive but

    Gadiraju, V., Kane, S., Dev, S., Taylor, A., Wang, D., Denton, E., & 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 (pp. 205 –216). ACM. https://doi.org/10.1145/3593013.3593989

    arXiv 2023

  22. [22]

    Block, L. (n.d.). Stereotypes about people with disabilities. Disability History Museum. Retrieved January 16, 2025, from https://www.disabilitymuseum.org/dhm/edu/essay.html?id=24

    2025

  23. [23]

    Heung, S., Phutane, M., Azenkot, S., Marathe, M., & Vashistha, A. (2022). Nothing micro about it: Examining ableist microaggressions on social media. In Proceedings of the 24th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS '22) (Article 27, pp. 1–14). ACM. https://doi.org/10.1145/3517428.3544801

    arXiv 2022

  24. [24]

    Friedman, C. (2019). Mapping ableism: A two-dimensional model of explicit and implicit disability attitudes. Canadian Journal of Disability Studies, 8(3), 95–120. https://doi.org/10.15353/cjds.v8i3.509

    work page doi:10.15353/cjds.v8i3.509 2019

  25. [25]

    Hutchinson, B., Prabhakaran, V., Denton, E., Webster, K., Zhong, Y., & Denuyl, S. (2020). Unintended machine learning biases as social barriers for persons with disabilities. SIGACCESS Accessibility and Computing, 125, 1–1. https://doi.org/10.1145/3386296.3386305

    arXiv 2020

  26. [26]

    Nakamura, K. (2019). My algorithms have determined you’re not human: AI-ML, reverse Turing-tests, and the disability experience. In Proceedings of the 21st International ACM SIGACCESS Conference on Computers and Accessibility (pp. 1–2). ACM. https://doi.org/10.1145/3308561.3353812

    arXiv 2019

  27. [27]

    Ahn, J., & Oh, A. (2021). Mitigating language-dependent ethnic bias in BERT. arXiv. https://doi.org/10.48550/arXiv.2109.05704

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2109.05704 2021

  28. [28]

    H., Sahay, S., Mazumder, S., Okur, E., Manuvinakurike, R., Beckage, N., Su, H., Lee, H., & Nachman, L

    Kumar, S. H., Sahay, S., Mazumder, S., Okur, E., Manuvinakurike, R., Beckage, N., Su, H., Lee, H., & Nachman, L. (2024). Decoding biases: Automated methods and LLM judges for gender bias detection in language models. arXiv. https://doi.org/10.48550/arXiv.2408.03907

    Pith/arXiv arXiv doi:10.48550/arxiv.2408.03907 2024

  29. [29]

    Hsu, T., & Thompson, S. A. (2023, February 8). Disinformation researchers raise alarms about A.I. chatbots. The New York Times . https://www.nytimes.com/2023/02/08/technology/ai-chatbots-disinformation.html

    2023

  30. [30]

    J., & Narayanan, A

    Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from language corpora contain human -like biases. arXiv. https://doi.org/10.48550/arXiv.1608.07187

    Pith/arXiv arXiv doi:10.48550/arxiv.1608.07187 2017

  31. [31]

    -W., Natarajan, P., & Peng, N

    Sheng, E., Chang, K. -W., Natarajan, P., & Peng, N. (2019). The woman worked as a babysitter: On biases in language generation. arXiv. https://doi.org/10.48550/arXiv.1909.01326

    Pith/arXiv arXiv doi:10.48550/arxiv.1909.01326 2019

  32. [32]

    Magee, L., Ghahremanlou, L., Soldatic, K., & Robertson, S. (2021). Intersectional bias in causal language models. arXiv. https://doi.org/10.48550/arXiv.2107.07691

    Pith/arXiv arXiv doi:10.48550/arxiv.2107.07691 2021

  33. [33]

    N., Srinath, M., & Wilson, S

    Venkit, P. N., Srinath, M., & Wilson, S. (2022). A study of implicit bias in pretrained language models against people with d isabilities. In Proceedings of the 29th International Conference on Computational Linguistics (pp. 1324 –1332). International Committee on Computational Linguistics. Retrieved January 16, 2025, from https://aclanthology.org/2022.co...

    2022

  34. [34]

    Beguš, N. (2024). Experimental narratives: A comparison of human crowdsourced storytelling and AI storytelling. Humanities and Social Sciences Communications, 11(1), 1392. https://doi.org/10.1057/s41599-024-03868-8

    doi:10.1057/s41599-024-03868-8 2024

  35. [35]

    Perfectly to a Tee: Understanding User Perceptions of Personalized LLM-Enhanced Narrative Interventions

    Bhattacharjee, A., Xu, S. Y., Rao, P., Zeng, Y., Meyerhoff, J., Ahmed, S. I., Mohr, D. C., Liut, M., Mariakakis, A., Kornfield, R., & Williams, J. J. (2024). “It explains what I am currently going through perfectly to a tee”: Understanding user perceptions on LLM-enhanced narrative interventions. arXiv. https://doi.org/10.48550/arXiv.2409.16732

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2409.16732 2024

  36. [36]

    Plaat, A., Wong, A., Verberne, S., Broekens, J., van Stein, N., & Back, T. (2024). Reasoning with large language models, a survey. arXiv. https://doi.org/10.48550/arXiv.2407.11511 16

    doi:10.48550/arxiv.2407.11511 2024

  37. [37]

    M., & Clavel, C

    Chhun, C., Suchanek, F. M., & Clavel, C. (2024). Do language models enjoy their own stories? Prompting large language models for automatic story evaluation. Transactions of the Association for Computational Linguistics, 12, 1122–1142. https://doi.org/10.1162/tacl_a_00689

    doi:10.1162/tacl_a_00689 2024

  38. [38]

    Huang, S., Mamidanna, S., Jangam, S., Zhou, Y., & Gilpin, L. H. (2023). Can large language models explain themselves? A study of LLM-generated self-explanations. arXiv. https://doi.org/10.48550/arXiv.2310.11207

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.11207 2023

  39. [39]

    Tian, Y., Huang, T., Liu, M., Jiang, D., Spangher, A., Chen, M., May, J., & Peng, N. (2024). Are large language models capable of generating human- level narratives? arXiv. https://doi.org/10.48550/arXiv.2407.13248

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.13248 2024

  40. [40]

    Chew, R., Bollenbacher, J., Wenger, M., Speer, J., & Kim, A. (2023). LLM-assisted content analysis: Using large language models to support deductive coding. arXiv. https://arxiv.org/abs/2306.14924

    Pith/arXiv arXiv 2023

  41. [41]

    Phutane, M., Seelam, A., & Vashistha, A. (2024). How toxicity classifiers and large language models respond to ableism. arXiv. https://doi.org/10.48550/ARXIV.2410.03448

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2410.03448 2024

  42. [42]

    Global World Down Syndrome Day. (2024). Assume that I can [Video]. YouTube. Retrieved January 16, 2025, from https://www.youtube.com/watch?v=92ivgabfdPQ

    2024

  43. [43]

    OpenAI. (n.d.). Prompt engineering. Retrieved January 16, 2025, from https://platform.openai.com/docs/guides/prompt-engineering

    2025

  44. [44]

    Emerson E, & Glover G. (2012). The ‘transition cliff’ in the administrative prevalence of intellectual disability. Tizard Lea rning Disability Review. 17(3):139–43. Doi: 10.1108/13595471211240988

    work page doi:10.1108/13595471211240988 2012

  45. [45]

    L., & Shogren, K

    Wehmeyer, M. L., & Shogren, K. A. (2016). Self -determination and choice. In N. N. Singh (Ed.), Handbook of evidence-based practices in intellectual and developmental disabilities (pp. 561–584). Springer International Publishing. https://doi.org/10.1007/978-3-319-26583-4_21

    work page doi:10.1007/978-3-319-26583-4_21 2016

  46. [46]

    Braddock, D., & Parish, S. (2001). An institutional history of disability. In G. L. Albrecht, K. D. Seelman, & M. Bury (Eds.), Handbook of disability studies (1st ed., pp. 11–68). Sage Publications

    2001

  47. [47]

    {}" """ Appendix 2: All Questions Asked for All LLMs Tested: Percentage of results that are “Yes

    Guo, Y., Guo, M., Su, J., Yang, Z., Zhu, M., Li, H., Qiu, M., & Liu, S. S. (2024). Bias in large language models: Origin, evaluation, and mitigation. arXiv. https://arxiv.org/abs/2411.10915 17 APPENDIX Appendix 1: Question Prompt: LLM Prompt used in story comprehension. """ What are the details of the person in the text in quotes? Respond with a json resp...

    Pith/arXiv arXiv 2024