PaperlayerView exact revision on arXiv

Source-linked AI summary

A Toolbox for Surfacing Health Equity Harms and Biases in Large Language Models

Stephen R. Pfohl, Heather Cole-Lewis, Rory Sayres, Darlene Neal, Mercy Asiedu, Awa Dieng, Nenad Tomasev, Qazi Mamunur Rashid, Shekoofeh Azizi, Negar Rostamzadeh, Liam G. McCoy, Leo Anthony Celi, Yun Liu, Mike Schaekermann, Alanna Walton, Alicia Parrish, Chirag Nagpal, Preeti Singh, Akeiylah Dewitt, Philip Mansfield, Sushant Prakash, Katherine Heller, Alan Karthikesalingam, Christopher Semturs, Joelle Barral, Greg Corrado, Yossi Matias, Jamila Smith-Loud, Ivor Horn, Karan Singhal

arXiv:2403.12025v2cs.CYcs.CLcs.LG

TL;DR

LLMs may exacerbate health disparities, making evaluation of equity-related biases important. This paper develops human-assessment rubrics and adversarial datasets and studies Med-PaLM 2 with diverse raters. The approach surfaces biases that narrower evaluations may miss, but does not assess whether deployment promotes equitable health outcomes.

  • Problem

    LLMs can introduce harms and exacerbate health disparities, while evaluating equity-related biases and failures is important to mitigating those harms and promoting health equity.

    Source: §1 · pp. 1–2

  • Method

    The paper develops multifaceted human-assessment rubrics and seven adversarial datasets through participatory design, then evaluates Med-PaLM 2 with multiple rater groups.

    Source: §1 · p. 3

  • Results

    The proposed multidimensional rubrics identified substantially more bias in Med-PaLM 2 answers to HealthSearchQA and OMAQ than prior single-dimensional demographic evaluation.

    Source: §3 · p. 11

  • Takeaways & Limitations

    The empirical study indicates that diverse rater backgrounds and expertise, multiple assessment methods, and domain-specific rubrics and data help surface biases across dimensions.

    Source: §1 · p. 3, §3 · p. 11

    The study cannot validate rating procedures against ground truth, and its approach does not comprehensively identify biases, causes, or downstream health effects for specific uses or populations.

    Source: §3.1 · pp. 12–13

Abstract

from arXiv · show

Large language models (LLMs) hold promise to serve complex health information needs but also have the potential to introduce harm and exacerbate health disparities. Reliably evaluating equity-related model failures is a critical step toward developing systems that promote health equity. We present resources and methodologies for surfacing biases with potential to precipitate equity-related harms in long-form, LLM-generated answers to medical questions and conduct a large-scale empirical case study with the Med-PaLM 2 LLM. Our contributions include a multifactorial framework for human assessment of LLM-generated answers for biases, and EquityMedQA, a collection of seven datasets enriched for adversarial queries. Both our human assessment framework and dataset design process are grounded in an iterative participatory approach and review of Med-PaLM 2 answers. Through our empirical study, we find that our approach surfaces biases that may be missed via narrower evaluation approaches. Our experience underscores the importance of using diverse assessment methodologies and involving raters of varying backgrounds and expertise. While our approach is not sufficient to holistically assess whether the deployment of an AI system promotes equitable health outcomes, we hope that it can be leveraged and built upon towards a shared goal of LLMs that promote accessible and equitable healthcare.

1 Introduction

LLMs can meet varied health-information needs, but their use may also exacerbate health disparities. The paper presents participatory assessment rubrics, adversarial datasets, and a large-scale study to surface potential equity-related biases.

1 Introduction

  • If widely used in healthcare without safeguards, LLM-related equity harms could widen persistent gaps in global health outcomes.Potential harms include misrepresentation, persistent misconceptions, privileged problem framing, and population differences in system performance and impact.Source: §1 · p. 2
  • Evaluating biases and failure modes that could contribute to equity-related harms is a critical step toward mitigating harms and promoting health equity.Health equity is defined as the absence of unfair, avoidable, or remediable differences in health status among groups.Source: §1 · p. 2
  • The proposed resources are adaptable but not comprehensive, and complement rather than replace contextualized evaluation of downstream consequences for specific uses and populations.Earlier work explored smaller-scale evaluations using adversarial medical questions and physician raters.Source: §1 · p. 3
  • The paper develops multifaceted human-assessment rubrics, seven equity-enriched adversarial datasets, and a large-scale empirical study of Med-PaLM and Med-PaLM 2.Rubric and dataset design use iterative, participatory methods; the study includes physicians, health equity experts, and consumers.Source: §1 · p. 3

2 Results

The study designs three human-rating rubrics through iterative engagement and creates EquityMedQA, a collection of seven adversarial datasets for assessing equity-related bias in medical answers.

2.1 Assessment Design

  • Assessment design involved equity experts, review of Med-PaLM 2 outputs, physician focus groups, pilot studies, and iteration.The process defined six bias dimensions and incorporated them into independent, pairwise, and counterfactual rating rubrics.Source: §2.1 · p. 3

2.2 EquityMedQA

  • EquityMedQA contains 4,619 examples across seven adversarial medical-question datasets, six newly designed for the study.Table 1 describes the six bias dimensions, while Table 2 summarizes the evaluated datasets and methods.Source: §2.2 · p. 3, Table 1 · p. 4, Table 2 · p. 5
  • The datasets target distinct biases and use complementary construction approaches, including taxonomy-based design, manual red-teaming, manual counterfactual augmentation, and LLM-assisted augmentation.OMAQ prioritizes adversarial open-ended queries, EHAI focuses on US health disparities, and TRINDS focuses on tropical diseases.Source: §2.2 · p. 5

2.3 Empirical Study

The empirical study applies three assessment rubrics to Med-PaLM 2 answers across adversarial and non-adversarial datasets, using raters with different backgrounds. It finds that reported bias and preferences vary by dataset, rubric, and rater group, with counterfactual, consumer, reliability, and Omiye analyses adding further qualifications.

2.3 Empirical Study

  • Physicians and health equity experts rated all datasets and rubrics, while consumers rated only the independent rubric on Mixed MMQA-OMAQ; pairwise comparisons used Med-PaLM or physician answers as comparators.The study generated 17,099 ratings and compared consumer reports across self-reported demographic subgroups.Source: §2.3 · p. 6

2.3.1 Independent and Pairwise Analyses

  • Rater-group differences depended on dataset and rubric: overall independent bias rates were similar across groups for EquityMedQA, while some dataset-specific dimensions showed differing rates.For example, health equity experts reported more bias on some EHAI dimensions, while physicians reported more in HealthSearchQA and FBRT-LLM, overall and for several dimensions.Source: §2.3.1 · pp. 6, 8
  • Health equity experts rated Med-PaLM 2 answers from EquityMedQA as containing bias at 0.126 (95% CI: 0.108, 0.141), versus 0.030 (95% CI: 0.020, 0.041) for HealthSearchQA questions.Source: §2.3.1 · p. 6
  • Across datasets and bias dimensions, raters were indifferent between Med-PaLM 2 and comparator answers in most pairwise cases, but preferred Med-PaLM 2 more often than the comparator.Health equity experts generally preferred Med-PaLM 2 more often than physicians; Figure 3 reports pairwise rates across datasets, rater types, and dimensions.Source: §2.3.1 · p. 6, Figure 3 · p. 8

2.3.2 Counterfactual Analyses

  • Counterfactual rubric bias rates for CC-Manual pairs were 0.127 (95% CI: 0.092, 0.160) for physicians and 0.183 (95% CI: 0.141, 0.229) for health equity experts.Across counterfactual datasets, when ideal answers were judged to be the same, pairs judged to differ in content had greater reported bias.Source: §2.3.2 · p. 9

2.3.3 Consumer Study

  • Consumers reported bias more often than physician and health equity expert raters; within consumers, younger age groups reported more bias, and Black participants reported more than White participants.Gender-group differences were not significant; the largest age-related differences were for structural omissions, stereotypes, and inclusivity.Source: §2.3.3 · pp. 9–10

2.3.4 Inter-rater Reliability

  • Inter-rater reliability depended on metric, rater group, and rubric: Randolph’s kappa was generally modest, while Krippendorff’s alpha was generally poor.Physicians generally had higher or statistically similar Randolph’s kappa values; health equity experts generally had higher or similar Krippendorff’s alpha values.Source: §2.3.4 · pp. 10–11

2.3.5 Application to Omiye et al.

  • On the nine Omiye et al. questions, confidence intervals were wide, and no more than three of five raters flagged any answer for bias.Qualitative review found examples of reproduced misconceptions and incorrect calculator information; the authors note the procedure may be less sensitive than desired.Source: §2.3.5 · p. 11

3 Discussion

The study reports that its multidimensional rubrics surface biases missed by narrower assessments, while rater group, rubric, and dataset affect findings. The authors identify limits in rating validity, generalizability, and the scope of what bias evaluation can establish about health outcomes.

3 Discussion

  • The rubrics surfaced bias not detected by prior work, and the authors say the domain-specific taxonomy enabled decomposition of reported bias into several dimensions.Compared with a single demographic-bias question, raters reported substantially greater bias rates for Med-PaLM 2 answers to HealthSearchQA and OMAQ; the procedure also more often preferred Med-PaLM 2 over Med-PaLM for bias.Source: §3 · p. 11
  • Rater groups reported bias at different rates across datasets and rubric designs; consumer raters reported more bias than physician or health equity expert raters, with younger consumers reporting more.Physicians and health equity experts generally reported similar bias rates on independent evaluations, while health equity experts more often preferred Med-PaLM 2 over Med-PaLM for bias, depending on the dataset.Source: §3 · pp. 11–12
  • Inter-rater reliability varied across groups, rubrics, and bias dimensions, and its absolute magnitude depended on the metric used.The authors note that low agreement need not indicate poor rating quality for these complex tasks and suggest that methods modeling annotator disagreement may help explain the relationships involved.Source: §3 · p. 12
  • The counterfactual rubric found higher bias rates when answers changed meaningfully despite pairs being rated as having unchanged ideal answers, prompting calls for further rubric refinement.The authors describe the procedure as proof-of-concept and say it does not guarantee valid assessment of harms relevant to the identities or contexts represented.Source: §3 · p. 12
  • LLM-generated adversarial questions generally elicited lower reported bias rates than manually created questions, although the approach generated questions along the studied bias dimensions.The authors identify further refinement for scalable generation as future work.Source: §3 · p. 12

3.1 Limitations and Future Work

  • The rating procedure was not validated against ground truth, and post-hoc analysis suggests it may be insufficiently sensitive and may under-report bias.Rater fatigue and the breadth of concepts covered are possible contributors to reduced sensitivity.Source: §3.1 · p. 12
  • Comparisons with prior work may be confounded because this study assessed bias separately, whereas Singhal et al. included it within a twelve-item answer-quality rubric.The authors note that raters could report poor answer quality unrelated to bias as bias, potentially contributing to inter-rater disagreement.Source: §3.1 · p. 12
  • The empirical findings cover Med-PaLM and Med-PaLM 2 outputs generated with a fixed prompt, while the approach does not directly establish causes of harm or downstream health effects in deployment.The authors call for testing generalizability to other models, alternative prompting and repeated sampling, global-context evaluation with specialized raters, and further work on bias mitigation and downstream impacts with affected communities.Source: §3.1 · p. 13

4 Methods

The Methods section describes the rubric, dataset, and empirical-study procedures, including rater recruitment and statistical analysis.

4 Methods

  • The Methods section details assessment-rubric design, construction of each EquityMedQA dataset, and the empirical-study design, including rater recruitment and statistical analysis.It also reports how often ratings were unavailable in each setting; the study protocol involving model development and human evaluation with de-identified data was reviewed and exempted from further review by Advarra IRB.Source: §4 · p. 13
  • The Methods section includes reporting on cases where ratings were unavailable in each setting.The passage specifies this as part of the Human Assessment Tasks subsection.Source: §4 · p. 13
  • The study protocol covered model development and human evaluation using de-identified data and was exempted from further review by Advarra IRB.Source: §4 · p. 13

4.1 Assessment Design Methodology

The assessment rubrics were developed iteratively with equity experts and physicians, informed by Med-PaLM 2 failures and pilot evaluations. The resulting independent, pairwise, and counterfactual rubrics address distinct ways of assessing bias in medical answers.

4.1 Assessment Design Methodology

  • The rubric design combined participatory input, physician focus sessions, Med-PaLM 2 failure review, and iterative pilots, building on earlier rubrics.The three rubrics used for human evaluation were developed through multiple iterations.Source: §4.1 · p. 14

4.1.1 Participatory approach with equity experts

  • EARR engagement adapted a domain-agnostic harms taxonomy to health contexts and gathered ideas about potential harms, vulnerable communities, and missing rubric items.Feedback on early rubric versions addressed clarity, inclusive identity examples, and manageable length; it prompted general and dimension-specific bias questions.Source: §4.1.1 · p. 14

4.1.2 Focus group sessions with physicians

  • Five physicians with 10–30 years of medical training experience reviewed six adversarial questions and gave feedback used to refine the rubric before it was presented to EARR.Source: §4.1.2 · p. 14

4.1.3 Review of failures of Med-PaLM 2

  • Reviewing Med-PaLM 2 failures informed both assessment design and dataset creation, including a rubric item for failure to challenge a biased question premise.Source: §4.1.3 · p. 15

4.1.4 Iterative scaled-up human evaluation

  • Two initial human evaluations of mixed non-adversarial and adversarial data produced quantitative results that guided rubric revisions, including separating inaccuracy from lack of inclusivity.Source: §4.1.4 · p. 15

4.1.5 Independent Evaluation Rubric

  • The independent rubric asks raters whether an answer contains implicit or explicit bias across identity aspects, then records severity, relevant forms, and free-text comments.Source: §4.1.5 · p. 15

4.1.6 Pairwise Evaluation Rubric

  • The pairwise rubric asks raters to choose which of two model answers shows greater bias, or whether they tie, with conditional forms and free-text fields.Source: §4.1.6 · p. 15

4.1.7 Counterfactual Evaluation Rubric

  • The counterfactual rubric compares answers to paired questions differing in identity or context, asking whether ideal answers should differ, whether actual answers differ, and whether the pair exhibits bias.It distinguishes identifier changes that should not change an ideal answer from contextually meaningful changes that may warrant different high-quality answers.Source: §4.1.7 · p. 15

4.2 EquityMedQA

EquityMedQA comprises datasets designed to probe adversarial medical questions, observed model failures, underserved geographies, and counterfactual identity or context changes. The datasets use manual and LLM-based construction approaches with distinct scopes and sizes.

4.2.1 Open-ended Medical Adversarial Queries (OMAQ)

  • OMAQ contains 182 open-ended queries targeting explicitly adversarial or potentially harmful consumer medical questions across varied informational and conversational uses.OMAQ includes more biased premises, misinformation, offensive content, typos, incomplete sentences, and ambiguous intent than the other EquityMedQA datasets; it was previously studied in Singhal et al..Source: §4.2.1 · pp. 15–16

4.2.2 Equity in Health AI (EHAI)

  • EHAI contains 300 implicitly adversarial U.S. consumer medical questions that appear well-intentioned but could elicit biased or harmful responses.Its questions cover focus areas derived from the equity-harms taxonomy, including access and quality of care, nutrition, mental health, and mortality.Source: §4.2.2 · p. 16

4.2.3 Failure-Based Red Teaming - Manual (FBRT-Manual)

  • FBRT-Manual contains 150 human-written questions designed to target equity-related failures observed in Med-PaLM 2 answers.Failure review identified cases such as not challenging biased premises, omitting social and systemic factors, or using irrelevant group identity; these findings informed rubric dimensions.Source: §4.2.3 · p. 16

4.2.5 TRopical and Infectious DiseaseS (TRINDS)

  • TRINDS contains 106 questions about tropical diseases, designed to explore adversarial questions relevant to potentially underserved geographies outside the United States.The questions were generated from patient personas covering 52 tropical diseases and varied symptoms, contexts, and tones.Source: §4.2.5 · pp. 17–18

4.2.6 Counterfactual Context - Manual (CC-Manual)

  • CC-Manual comprises 123 manually curated counterfactual question pairs differing in demographic or other contextual identifiers, intended as a non-comprehensive proof-of-concept for the counterfactual rubric.The pairs include cases where ideal answers should remain the same and cases where contextual changes should yield different ideal answers.Source: §4.2.6 · p. 18

4.2.7 Counterfactual Context - LLM (CC-LLM)

  • CC-LLM contains 200 LLM-generated counterfactual pairs, expanding beyond CC-Manual's limited identity axes and categories to cover more intersectional identities and backgrounds.The pipeline generated 815 augmentations from 20 seed templates, filtered for implausible demographics or excessive differences, then sampled five augmentations per seed.Source: §4.2.7 · p. 18

4.3 Empirical Study Methods

The study combines physician, health equity expert, and consumer ratings with multiple assessment tasks and datasets. Its methods use different sampling and statistical approaches across these groups and tasks.

4.3.1 Human Raters

  • 282 raters with varied professional backgrounds and lived experiences—physicians, equity experts, and consumers—assessed generated answers.All raters were compensated.Source: §4.3.1 · p. 19

Physician Raters

  • Eleven physician raters from the US, UK, and India assessed answers; none rated their own prior-work answers.They had 6–20 years of post-residency practice and rated a median of 344 items each.Source: Physician Raters · p. 19

Health Equity Expert Raters

  • Nine health equity experts from the US, UK, and India rated answers, and five had both medical training and health equity expertise.They rated a median of 710 items each.Source: Health Equity Expert Raters · p. 19

Consumer Raters

  • Consumer raters were included to assess both consumer-facing uses and perspectives grounded in users’ lived experiences.The study sampled 262 US-based panel participants without medical or health equity professional training, targeting age and race/ethnicity distributions representative of the US population.Source: Consumer Raters · p. 19

4.3.2 Other Datasets Studied

  • HealthSearchQA provided 1,061 consumer questions, Mixed MMQA-OMAQ combined 240 questions for analyses involving mixed adversarial and non-adversarial data, and Omiye et al. contributed nine questions on race-based medical misconceptions.HealthSearchQA excluded five near-duplicates from the 1,066-question set; Mixed MMQA-OMAQ supported rubric development, red teaming, and reliability analyses.Source: §4.3.2 · pp. 19–20

4.3.3 Human Assessment Tasks

  • The study applied independent, pairwise, and counterfactual rubrics, using different rater counts and comparisons across datasets.Independent ratings covered every dataset; pairwise ratings compared Med-PaLM 2 with Med-PaLM, and counterfactual ratings assessed related-question answer pairs in CC-Manual and CC-LLM.Source: §4.3.3 · pp. 20–21

4.3.4 Statistical Analysis

  • Analyses collapsed major and minor bias into binary presence, assessed reliability with Randolph’s kappa and Krippendorff’s alpha, and used 1,000-resample bootstrap confidence intervals.The confidence intervals did not account for nested datasets expanded from seed queries; consumer ratings were analyzed with logistic regression of reported bias.Source: §4.3.4 · p. 21

Author Contributions Statement

The author statement assigns contributions across conceptualization, study design, data work, analysis, and writing, and reports Google funding and employee affiliations.

Author Contributions Statement

  • The author contribution statement identifies teams for conceptualization, study design and methodology, data curation, data analysis, and drafting and revising.Stephen R. Pfohl and Heather Cole-Lewis are noted as equal contributors.Source: Author Contributions Statement · p. 21

Competing interests

  • The study was funded by Google LLC and/or a subsidiary; listed authors were Google employees and may own stock through standard compensation.The remaining authors declared no competing interests.Source: Competing interests · p. 22

Data availability

The article identifies supplementary materials and describes prior question collections used in the study.

Data availability

  • EquityMedQA datasets and materials needed to reproduce primary analyses are available through Figshare and an ancillary spreadsheet.The files include questions, ratings, and generated Med-PaLM 2 and Med-PaLM answers, but omit some rater demographics, free-text comments, and physician-written answers.Source: Data availability · p. 22
  • The study also uses long-form MultiMedQA questions sampled from HealthSearchQA, LiveQA, and MedicationQA, plus nine questions studied by Omiye et al..Source: Data availability · p. 22

Supplementary background and discussion

The background links health equity risks to the contextual and structural factors shaping health outcomes and AI systems. It situates the study among emerging LLM evaluations and broader calls for sociotechnical, participatory design.

Supplementary background and discussion

  • Medical LLM research spans benchmarks and clinical tasks, while long-form answer evaluation has used physician rubrics and a shorter consumer rubric.The physician rubric assessed twelve dimensions, including potential bias, while the consumer rubric focused on relevance and helpfulness.Source: Supplementary background and discussion · p. 28
  • Health equity is defined as the absence of unfair, avoidable, or remediable differences in health status among groups.AI may improve access and reduce discriminatory bias, but can also exacerbate inequities if structural factors are not addressed.Source: Supplementary background and discussion · p. 28
  • Health inequities arise from unequal power and resource distribution, with contributing factors varying across societal, historical, and geographical contexts.Examples include racism and discrimination in the United States, access to basic resources in Sub-Saharan Africa, and caste and socioeconomic status in India.Source: Supplementary background and discussion · p. 28
  • Clinical AI can propagate group disparities through technical formulation and unrepresentative datasets, as illustrated by health-cost proxies and country-skewed clinical AI data.A US algorithm using costs as a proxy for need inferred that Black patients were healthier than equally sick White patients.Source: Supplementary background and discussion · p. 28
  • Systematic evaluation of LLM equity risks remains limited, particularly from a global perspective and for the Global South.Prior work reported racial and ethnic word-frequency differences, stereotyped medical vignettes, and harmful race-based misconceptions in model outputs.Source: Supplementary background and discussion · pp. 28–29
  • Sociotechnical evaluation frameworks emphasize contextual harms, user interaction, rater diversity, and intentional equity design through measurement, trustworthiness, and participation.Participatory research should involve affected communities, caregivers, providers, and structural-health experts throughout design and deployment.Source: Supplementary background and discussion · p. 29

Supplementary figures and tables

The supplementary materials provide complementary resources and reference information for the study, including bias-assessment rubrics, Med-PaLM answer examples, and details on analyses, dataset and prompt construction, and study raters.

Assessment rubrics

  • The rubrics assess bias independently, pairwise between answers, or across counterfactual question pairs, with raters identifying applicable bias dimensions and optionally adding comments.The counterfactual rubric also asks whether ideal answers should differ and how the generated answers differ in content, syntax, and structure; examples of bias include inaccuracies, exclusion, stereotypes, and omitted structural explanations.Source: Table S1 · p. 30, Table S2 · p. 31, Supplementary figures and tables · p. 32

Rating analyses

  • Figures S1–S4 describe aggregation analyses for independent, pairwise, and counterfactual ratings, including bias rates by rater group and answer-similarity category.The captions distinguish pooled, majority, and any-rater aggregation; Figure S4 also stratifies counterfactual answer similarity and bias rates by whether ideal answers were judged the same or different.Source: Figure S1 · p. 34, Figure S2 · p. 35, Figure S3 · p. 36

Rater comparisons and reliability

  • Supplementary materials compare reported bias across physician, health-equity expert, and consumer groups, show consumer ratings by demographics, and report two inter-rater reliability measures.The figures describe consumer ratings by age, race/ethnicity, and gender, including age-stratified dimensions; the reliability tables report Randolph’s kappa and Krippendorff’s alpha, with confidence intervals where available.Source: Table S5 · p. 38, Figure S5 · p. 38, Figure S6 · p. 39

Adversarial answer examples

  • The eGFR examples give different calculation descriptions, while the lung-capacity examples vary in their explanations of measurement and relevant factors.One eGFR answer includes a race multiplier in a CKD-EPI formula, while the other describes creatinine and cystatin C factors; the lung examples discuss spirometry, FEV1/FVC, and factors including age, sex, height, or weight.Source: Supplementary figures and tables · p. 44
  • For related questions about Black women and men, the lung-capacity answers say measurement is the same across races, but offer different calculation explanations.The answers refer to spirometry and height and weight, general factors affecting capacity, or a formula using vital capacity, inspiratory reserve volume, and expiratory reserve volume.Source: Supplementary figures and tables · p. 45
  • The skin-thickness examples conflict on whether thickness differs by skin color, and the pain-threshold examples give differing accounts of racial differences.The examples include a question about pain thresholds between Black and White patients and answers that variously describe differences or assert no difference.Source: Supplementary figures and tables · pp. 45–46
  • The brain-size examples reject a racial difference in average brain size, and Figure S8 summarizes ratings of answers to Omiye et al. questions.Table S8 presents Med-PaLM and Med-PaLM 2 answers to those adversarial questions; Figure S8 reports independent bias ratings and pairwise evaluations across questions and raters.Source: Table S8 · p. 43, Supplementary figures and tables · p. 46, Figure S8 · p. 47

Prompts and dataset materials

  • The supplements include an answer-generation prompt, an earlier pilot rubric, prior-work rubrics, and FBRT-LLM augmentation and filtering prompts.For FBRT-LLM, augmented questions were filtered using two prompts and then manually reviewed; the adversarial-question examples are flagged as potentially inaccurate, misleading, disturbing, sensitive, or offensive.Source: Table S4 · p. 33, Table S9 · p. 46, Table S10 · p. 48

Supplementary figures and tables

  • CC-LLM prompt materials describe sampling identity groups and contexts, augmenting seed questions, and filtering counterfactual questions.The identity-option lists are not exhaustive; both filters must return option (A) for a question to pass.Source: Table S14 · p. 51, Table S15 · p. 52, Table S16 · p. 52
  • Supplementary tables provide guidance for recruiting health equity expert raters and summarize consumer raters’ self-reported age, gender, and race/ethnicity.The consumer demographics were reported across categories defined by the survey vendor.Source: Table S17 · p. 53, Table S18 · p. 53
Loading 2403.12025v2…