Understanding Down Syndrome Stereotypes in LLM-Based Personas

Applications, bias, and safety2026ACMApproved editorial review

Authors: Chantelle Wu, Mengxu Pan, Peinan Wang, Nafi Nibras, Meida Li, Dajun Yuan, Zhixiao Wang, Jiahuan He, Mona Ali, Mirjana Prpa

Keywords: Persona, Large Language Models, Down Syndrome, Stereotype

Source: Open primary source (opens in a new tab)

10
Authors
8
Findings
14
Limitations
3
Evidence

Editorial summary

English

The final DIS 2026 paper investigates where stereotypes emerge when Persona-L generates interactive personas of people with Down syndrome using an LLM and RAG. Its strongest contribution is not evidence that a classifier solved bias, but a multi-layer account of representational failure: source data, interface parameters, and conversational delivery can each encode assumptions. Ten family caregivers, educators, and support professionals found some retrieved content factually plausible while judging the speech too long, polished, formulaic, and consistently cheerful. They also interpreted the age cap of 40, occupation options, and an overly positive abilities framework as assumptions introduced by the interface itself.

The version of record materially differs from the 2025 preprint. It adds Mengxu Pan, appears at DIS '26 under DOI 10.1145/3800645.3812894, and reframes the detector as a design probe for reflection rather than an objective arbiter. The final paper explicitly describes a qualitative design, adds thematic-analysis methods following Braun and Clarke's six phases, removes the preprint experiment associating RAG with a statistical reduction in stereotypes, and expands the ethical limitations. This review uses the final paper; the 35-page preprint is retained only for version lineage and removed appendices.

Persona-L is a prior web interface that generates profiles, presents key abilities, and supports persona chat. This paper adds a post-LLM module that segments each output into sentences, classifies them as neutral or as gender, profession, or Down-syndrome stereotypes, and displays a confidence score plus a GPT-4-generated explanation. It does not automatically rewrite text. This design fits the final research aim: disagreements between the detector and participant experience are intended to expose assumptions and prompt discussion about authenticity and representation.

The detector starts from a simplified Multi-Grain Stereotype Dataset derived from StereoSet and CrowS-Pairs. For the new class, the authors collected 30 seed sentences about family, education, and employment from myths-and-truths resources. Three researchers annotated them, reporting Fleiss' kappa 0.775, and GPT-4 expanded the set with 500 sentences; only an unspecified subset was systematically reviewed. ELECTRA, RoBERTa, DeBERTa, and DistilBERT were fine-tuned for three epochs with AdamW, cross-entropy, batch size 16, and early stopping. DistilBERT was excluded after 68% accuracy and the other three formed an ensemble.

The paper reports 16,139 training and 1,769 test instances, but its listed test classes sum to 1,795: 896 neutral, 647 profession, 218 gender, and 34 Down-syndrome examples. This 26-item discrepancy prevents reconstruction of the split. The ensemble reaches 82.06% accuracy, 82.16% F1, and 99% F1 on the Down-syndrome class, but the last result rests on only 34 reported items, largely downstream of 30 GPT-4-expanded seeds. Seed-family grouping, deduplication and leakage controls, exact checkpoints, learning rate, seeds, intervals, and validation split are not documented. The final paper appropriately acknowledges that the 99% result is fragile, susceptible to sampling variance and synthetic-data leakage, and not a production-ready benchmark.

The main study is qualitative. Ten adults with at least one year of relevant experience participated: family caregivers, teachers, speech-language pathologists, an occupational therapist, and other support roles. Nine were women and one was a man; seven lived in Canada, two in the United States, and one in Australia. Sessions lasted 60-80 minutes. Each person created three personas, explored one, answered six five-point Likert items, joined a semi-structured interview, and compared independent review with detector-assisted review. No person with Down syndrome participated directly.

The final analysis says it followed Braun and Clarke's six phases through repeated familiarization, semantic and latent coding, iterative grouping, and theme review. It reports four stereotype families, ableist devaluation, appearance assumptions, behavioural expectations, and intellectual capacity, plus cross-cutting themes concerning complexity, delivery authenticity, and interface design. This is more auditable than the preprint, but the paper does not state how many people coded, whether coding was independent, how disagreements were resolved, whether interviews were transcribed verbatim, or provide a codebook, audit trail, reflexivity account, or saturation criterion.

Participants emphasized heterogeneity in cognition, speech, health, behaviour, learning, and support needs, questioning whether any static persona could represent this diversity. The age cap of 40 was read as an echo of outdated medical expectations; occupation categories appeared either too aspirational or too restrictive; and the abilities framework risked idealization by omitting difficulty, support, and contextual variation. The useful UI/UX finding is that an apparently neutral interface can encode a social theory before the model produces any text.

Caregivers and professionals recognized plausible advice such as visual schedules, checklists, and educational support, but separated factual correctness from authenticity. Generated speech was often excessively long, sophisticated, and structured as answer, explanation, and follow-up question. Some considered it inconsistent with age, development, or embodied communication constraints. The uniformly enthusiastic tone reproduced the stereotype that people with Down syndrome are always happy, erasing anger, sadness, anxiety, and adulthood. These observations justify examining length, tone, emotional variability, and context; they do not authorize one supposedly correct way for people with Down syndrome to speak.

Validity remains limited. Evidence that RAG improved factual accuracy comes from judgments in this small sample, not independent fact checking, retrieval precision, or citation-fidelity evaluation. Six Likert items were administered, but no means, distributions, uncertainty, reliability, or respondent-level outcomes are reported. No participant-level precision, recall, or agreement is given for the validation task. The survey and detector therefore do not demonstrate quantitative effectiveness.

Caregivers and specialists contribute important contextual knowledge but cannot substitute for the experience of the people represented. Their judgments are also filtered through relationships, accommodations, and potentially ableist assumptions; some quotations use low-expectation or “high functioning” framings that require critical context rather than adoption as design truth. The final paper acknowledges this tension, advises against replacing direct research, and explicitly excludes hiring, benefits, and medical decisions as appropriate uses. It proposes a second phase with direct participation, supported consent, and community governance, but that phase is not yet evidence.

Public reproducibility is insufficient. The final text and preprint sources are open, but this audit did not recover Persona-L code, the 80 stories and full provenance, 30 seeds, 500 expansions, data splits, checkpoints, prompts, parameters, outputs, survey responses, validation judgments, transcripts, codebook, or an executable environment. Consent and IRB approval are stated without a visible protocol identifier. The responsible conclusion is a useful qualitative one: stereotypes may emerge from a pipeline and interaction design even when content sounds correct, and a detector can open reflection. The paper does not establish authentic representation, general stereotype reduction, robustness of the 99% F1, design trustworthiness, or a substitute for direct participation.

Español

El artículo final de DIS 2026 estudia dónde aparecen los estereotipos cuando Persona-L genera personas interactivas de gente con síndrome de Down mediante un LLM y RAG. Su aportación más sólida no es demostrar que un clasificador haya resuelto el sesgo, sino mostrar que la representación puede fallar en varias capas: los datos que alimentan el sistema, los parámetros de la interfaz y la forma conversacional de la respuesta. Diez familiares, docentes y profesionales de apoyo consideraron que parte del contenido recuperado era factualmente plausible, pero que el habla seguía pareciendo demasiado larga, pulida, formulaica y alegre. También interpretaron el límite de edad de 40 años, ciertas opciones laborales y un marco de capacidades excesivamente positivo como supuestos incorporados por el propio diseño.

La versión de referencia difiere materialmente del preprint de 2025. Añade a Mengxu Pan, se publica en DIS '26 con DOI 10.1145/3800645.3812894 y reformula el detector como una sonda de diseño para provocar reflexión, no como árbitro objetivo. La versión final explicita un diseño cualitativo, añade análisis temático siguiendo las seis fases de Braun y Clarke, elimina el experimento del preprint que asociaba RAG con una reducción estadística de estereotipos y amplía las limitaciones éticas. Esta ficha usa la versión final; el preprint de 35 páginas solo se conserva para trazar los cambios y revisar anexos retirados.

Persona-L es una interfaz web previa que genera perfiles, presenta capacidades clave y permite conversar con la persona. El trabajo la extiende con un módulo posterior al LLM: separa cada salida en oraciones, las clasifica como neutral o como estereotipo de género, profesión o síndrome de Down y muestra confianza más una explicación generada por GPT-4. El sistema no corrige automáticamente el texto. Esta elección encaja con el objetivo final del artículo: hacer visibles desacuerdos entre el detector y la experiencia de los participantes y usarlos como punto de partida para discutir autenticidad y representación.

El detector parte de una versión simplificada del Multi-Grain Stereotype Dataset, derivado de StereoSet y CrowS-Pairs. Para la nueva clase, los autores reunieron 30 frases semilla sobre familia, educación y empleo a partir de materiales de mitos y verdades. Tres investigadores las anotaron, con kappa de Fleiss 0,775, y GPT-4 amplió el conjunto con 500 frases; solo se revisó sistemáticamente un subconjunto cuyo tamaño no se informa. ELECTRA, RoBERTa, DeBERTa y DistilBERT se ajustaron durante tres épocas con AdamW, cross-entropy, lote 16 y early stopping. DistilBERT quedó fuera por una exactitud del 68 % y los otros tres formaron un ensemble.

El artículo informa 16.139 ejemplos de entrenamiento y 1.769 de test, pero las clases de test publicadas suman 1.795: 896 neutrales, 647 de profesión, 218 de género y 34 de síndrome de Down. Esa discrepancia de 26 elementos impide reconstruir el split. El ensemble alcanza 82,06 % de exactitud, F1 82,16 % y F1 99 % en la clase de síndrome de Down, pero ese último valor descansa en solo 34 ejemplos reportados, generados en gran parte a partir de 30 semillas mediante GPT-4. No se documentan agrupación por familia de semilla, deduplicación o control de fuga, checkpoints exactos, learning rate, seeds, intervalos ni conjunto de validación. La propia versión final reconoce acertadamente que el 99 % es frágil, susceptible a varianza muestral y fuga de datos sintéticos, y no constituye un benchmark listo para producción.

El estudio principal es cualitativo. Participaron diez personas adultas con al menos un año de experiencia: familiares, docentes, logopedas, una terapeuta ocupacional y otros perfiles de apoyo. Nueve eran mujeres y uno hombre; siete residían en Canadá, dos en Estados Unidos y uno en Australia. Cada sesión duró 60-80 minutos. Cada participante creó tres personas, exploró una, contestó seis ítems Likert de cinco puntos, realizó una entrevista semiestructurada y comparó una revisión independiente con las alertas del detector. Ninguna persona con síndrome de Down participó directamente.

El análisis final dice haber seguido las seis fases de Braun y Clarke, con familiarización repetida, codificación semántica y latente, agrupación iterativa y revisión de temas. Identifica cuatro familias de estereotipos, devaluación capacitista, supuestos sobre apariencia, expectativas conductuales y capacidad intelectual, además de temas transversales sobre complejidad, autenticidad de la entrega e interfaz. La descripción es más auditable que la del preprint, pero no informa cuántas personas codificaron, si lo hicieron independientemente, cómo resolvieron desacuerdos, si hubo transcripción literal, ni publica codebook, audit trail, reflexividad o criterio de saturación.

Los participantes insistieron en la heterogeneidad de cognición, habla, salud, conducta, aprendizaje y apoyos. Cuestionaron que una persona estática represente esa diversidad. La limitación de edad a 40 años fue entendida como eco de expectativas médicas obsoletas; las ocupaciones parecían a veces demasiado aspiracionales y otras demasiado restrictivas; y el marco de capacidades podía idealizar a la persona al omitir frustraciones, necesidades y variación contextual. El hallazgo útil para UI/UX es que una interfaz aparentemente neutral puede codificar una teoría social antes de que el modelo genere una sola palabra.

Sobre las respuestas, familiares y profesionales reconocieron consejos plausibles, horarios visuales, listas, apoyos educativos, pero separaron exactitud factual de autenticidad. El habla generada era a menudo excesivamente extensa, sofisticada y estructurada como respuesta, explicación y pregunta de seguimiento. Algunas personas la consideraron incompatible con edad, desarrollo o condiciones físicas de comunicación. El tono constantemente entusiasta reforzaba el estereotipo de que la gente con síndrome de Down está siempre feliz, borrando enfado, tristeza, ansiedad y madurez. Estas observaciones apoyan revisar longitud, tono, variabilidad emocional y contexto; no autorizan a fijar una única forma correcta de hablar.

La validez tiene límites importantes. La evidencia de que RAG mejoró la exactitud procede de juicios de este pequeño grupo, no de verificación independiente de hechos, precisión de retrieval o fidelidad de citas. El artículo administra seis preguntas Likert pero no publica medias, distribuciones, incertidumbre, fiabilidad ni resultados individuales. Tampoco informa precisión, recall o acuerdo humano-sistema en la tarea de validación. Por tanto, el cuestionario y el detector no demuestran eficacia cuantitativa.

La elección de cuidadores y especialistas aporta conocimiento contextual, pero no sustituye la experiencia de las personas representadas. Sus opiniones también están filtradas por relaciones, adaptaciones e ideas capacitistas; algunas citas del artículo usan formulaciones de bajo nivel de expectativa o categorías como «high functioning» que necesitan contextualización crítica y no deben convertirse en verdad de diseño. La versión final reconoce esta tensión, desaconseja sustituir investigación directa y excluye explícitamente empleo, prestaciones y decisiones médicas como usos adecuados. Propone una segunda fase con participación directa, apoyos de consentimiento y gobernanza comunitaria, pero esa fase aún no forma parte de la evidencia.

La reproducibilidad pública es insuficiente. El texto final y el preprint con sus fuentes son accesibles, pero no se recuperaron el código de Persona-L, las 80 historias y su procedencia detallada, las 30 semillas, las 500 expansiones, los splits, checkpoints, prompts, parámetros, outputs, respuestas del cuestionario, juicios de validación, transcripciones, codebook ni un entorno ejecutable. El estudio afirma consentimiento e IRB, sin mostrar identificador de protocolo. La lectura responsable conserva un resultado cualitativo útil: los estereotipos pueden emerger del pipeline y del diseño de interacción incluso cuando el contenido parece correcto, y un detector puede servir para abrir una conversación. No demuestra representación auténtica, reducción general de sesgo, robustez del F1 99 %, confianza para diseño ni capacidad para reemplazar participación directa.

Research question

How do caregivers and specialists perceive stereotypes in personas of people with Down syndrome generated by LLMs, and at which points of the pipeline, data, interface, or conversational delivery, do they identify their appearance?

Method

Qualitative study with Persona-L and a stereotype detector used as a design probe. Ten participants created three personas, explored one, responded to six Likert items, conducted a semi-structured interview, and compared independent and assisted revision. The detector combines ELECTRA, RoBERTa, and DeBERTa fine-tuned on simplified MGSD plus synthetic Down syndrome data; the final text reports thematic analysis according to Braun and Clarke.

Sample: N=10: nine women and one man; seven residing in Canada, two in the United States, and one in Australia; family members, teachers, speech-language therapists, and other support professionals with at least one year of experience. No people with Down syndrome participated directly.

Findings

  • Stereotypes appear in data, interface parameters, and conversational form, not only in factually biased statements.
  • The perceived factuality of RAG content did not guarantee the authenticity of the persona.
  • Speech was judged too long, polished, formulaic, and poorly adjusted to age, development, or physical communication conditions.
  • The always-cheerful tone flattened emotional variability and reinforced infantilization.
  • The age limit of 40 years, occupations, and the capabilities framework encoded possible stereotypic assumptions in the UI.
  • The heterogeneity of cognition, communication, health, behavior, learning, and supports challenges any static persona.
  • The ensemble reports 82.06% accuracy and F1 82.16%, but the 99% F1 for Down syndrome uses only 34 reported examples.
  • The final version removes the quantitative inference from the preprint that RAG reduces stereotypes and positions the detector as a reflective probe.

Limitations

  • No person with Down syndrome participates directly.
  • Caregivers are imperfect proxies and may reproduce ableist or low-expectation assumptions.
  • Sample of ten people, nine women, predominantly Canadian and only in English.
  • The six Likert items lack published numerical results, uncertainty, and reliability.
  • Precision, recall, or human-system agreement in validation is not reported.
  • The thematic analysis does not document the number of coders, independence, disagreements, codebook, audit trail, reflexivity, or saturation.
  • The test is declared to have 1,769 items, but its four classes sum to 1,795.
  • The Down syndrome class has only 34 reported examples and depends on synthetic expansion from 30 seeds.
  • Split, deduplication, grouping by seeds, or leakage controls are not explained.
  • Checkpoints, learning rate, seeds, validation, intervals, and ensemble specification are missing.
  • Generator model, prompts, parameters, chunking, embeddings, retrieval, and Persona-L logs are not fully documented.
  • RAG factuality is based on participant perception, without external verification or citation audit.
  • IRB is declared without a visible identifier.
  • Code, data, checkpoints, instruments, responses, transcripts, or a reproducible environment were not publicly recovered.

What the study does not establish

  • It does not demonstrate that Persona-L authentically represents the experience or speech of people with Down syndrome.
  • It does not demonstrate that caregivers can substitute direct participation.
  • It does not demonstrate that RAG reduces stereotypes in general or guarantees correct facts.
  • It does not demonstrate that the 99% F1 is robust, leakage-free, or production-ready.
  • It does not demonstrate through the questionnaire accuracy, confidence, or utility for design.
  • It does not demonstrate that the detector captures all cultural stereotypes or explains their cause.
  • It does not validate the system for employment, benefits, medicine, or other sensitive decisions.
  • It does not allow reproducing the metrics or the qualitative analysis with the located public artifacts.

Traceability

Scope: Full text

Version: DIS '26 version of record, DOI 10.1145/3800645.3812894, pages 2371-2391; arXiv:2512.02275v1 retained for version comparison

Consulted source: https://dl.acm.org/doi/10.1145/3800645.3812894

Review: Codex full-text version-of-record, 35-page preprint visual, arXiv-source, version-lineage, qualitative-method, participatory-validity, dataset-arithmetic, synthetic-leakage, UI, ethics and public-artifact audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Persona-L LLM no identificado con precisión
  • GPT-4 para expansión sintética y explicaciones
  • ELECTRA
  • RoBERTa
  • DeBERTa
  • DistilBERT

Instruments and metrics

  • Persona-L: generación, capacidades clave y chat
  • Detector de cuatro clases con confianza y explicación
  • Seis ítems Likert de cinco puntos
  • Entrevista semiestructurada
  • Revisión independiente y asistida de estereotipos
  • Análisis temático de Braun y Clarke

Data used

  • Multi-Grain Stereotype Dataset simplificado
  • StereoSet
  • CrowS-Pairs
  • 30 frases semilla sobre síndrome de Down
  • 500 frases sintéticas generadas con GPT-4
  • 80 historias públicas curadas para RAG de Persona-L

Evidence and location

  • Metadata, full text, final method, results, limitations, annexes, and disclosure: ACM DIS '26 version of record, DOI 10.1145/3800645.3812894, pp. 2371-2391
  • Previous lineage, removed details, and extended appendices: arXiv:2512.02275v1, 35 pages
  • Audit of participation, dataset, leakage, metrics, UI, ethics, and artifacts: reports/verification/article-234-down-syndrome-persona-participatory-dataset-leakage-metric-and-artifact-audit.json