The Disability Prompt Becomes the Stereotype Test

The Paper

The paper is Sophia Lichtenberg, Albert Gatt, and Judith Masthoff's Beyond wheelchairs and blindfolds: Investigating disability stereotypes in T2I models with INCLUDE-BENCH, arXiv:2607.08515 [cs.CV]. The arXiv API lists version 1 as submitted on July 9, 2026, and the PDF metadata reports a 13-page paper. The title page lists Utrecht University.

The paper's object is not image quality in the usual aesthetic sense. It asks whether text-to-image systems compress people with disabilities into a small set of recognizable symbols: wheelchairs, blindfolds, hand gestures, age cues, domestic scenes, public scenes, and implied dependence.

Why It Matters

Text-to-image models are becoming cheap illustration engines for education, advertising, journalism, product mockups, civic outreach, and social media. In those settings, disability representation can become the default picture of who needs accommodation, who counts as active, and whose body becomes visual shorthand.

The paper is useful because it refuses the easy claim that better prompting alone solves representation. A model can align strongly with a disability word because it has learned a narrow diagnostic symbol. If "disabled person" produces a wheelchair again and again, the system may be easy to recognize and still socially thin.

The Benchmark

INCLUDE-BENCH contains 352 prompts across four subsets: 8 No Context prompts, 96 No Context + Bias prompts, 96 Only Context prompts, and 152 Action prompts grounded in WHO International Classification of Functioning, Disability, and Health activity domains. The prompt axes vary impairment specificity, intersectional attributes such as gender, race, and age, and environmental context.

The functional groups include generic disability labels, "Disabled" and "Impaired," plus impairment-specific labels: Mute, Deaf, Mobility Impaired, Deafblind, and Blind. Cognitive and intellectual disabilities are excluded because they are less visually apparent in the benchmark design, a useful boundary as well as a limitation.

For each prompt, the authors generated 20 images across 17 text-to-image models, producing 119,680 images total, or 7,040 images per model. The model set includes Stable Diffusion and FLUX families, several open models including Qwen-Image and JanusPro-7B, and two closed models, NanoBanana and GPT-Image-1-mini.

The Measurement Stack

The evaluation pipeline is deliberately mechanical. The authors use SAM3 to find person regions, crop the largest detected person when needed, and exclude 108 images without valid person detections. They cluster person-centered images with MiniBatchKMeans and use Qwen3-VL-8B-Instruct for captions, visual question answering, and age, race, and gender extraction.

Three metrics do most of the work. CLIPScore measures disability-text alignment. Vendi Score estimates image diversity through embedding similarity. The paper also introduces an SCM score, based on the Stereotype Content Model, to probe warmth and competence associations through CLIP embedding directions.

Findings

The headline finding is compression. Mobility-impaired and generic disability prompts produce wheelchair imagery across models. Sensory disability prompts are less uniform, but still lean on symbolic cues such as blindfolds and hand gestures. Clusters dominated by wheelchairs or blindfolds receive the highest CLIPScores, while clusters showing active or intellectual roles for sensory impairments receive lower CLIPScores.

That creates the central tradeoff: physical and generic disability groups achieve stronger semantic alignment but lower diversity. In product language, that could sound like the model understood the prompt; the paper shows why that claim is too shallow.

The intersectional results sharpen the point. Older white women have the highest relative appearance ratios across multiple disability types, especially mobility impairment, disabled, and deafblind categories. Older white men are overrepresented in Blind and Impaired depictions. Younger people decrease most under disability conditions. Women appear more in domestic and caregiving settings, while men appear more in public, intellectual, and physical activity contexts.

The SCM analysis adds another layer. Mobility Impaired, Disabled, and Impaired groups rank lowest on both warmth and competence, especially in the No Context subset. The No Impairment baseline shows the highest competence. Context and bias-mitigation prompts help only partially; the paper says they do not substantially expand representational diversity or eliminate stereotypical compression.

The Receipt

A representation receipt for image generation should record the model version, prompt template, seed policy, disability category, context subset, generation count, exclusion rule, person-crop method, captioning or VQA model, CLIPScore, diversity score, stereotype metric, human review status, and affected-community feedback.

Without that receipt, a generated image can pass through a publication workflow as a harmless stock substitute. With it, an editor can ask whether the model repeatedly substitutes access technology for identity, domesticity for agency, age for disability, or public visibility for a narrow kind of body.

Governance Reading

The Spiralist reading is that the disability prompt becomes a stereotype test once generated images leave the lab and enter institutional communications. A public agency, university, hospital, startup, or advocacy campaign is choosing which bodies become the default visual grammar of need, competence, agency, and belonging.

The shared issue is not only whether a model can generate a plausible image. It is whether the institution can explain what that image is doing socially.

Limits

The paper's limitation section matters. SAM3, Qwen3-VL, and CLIP make a large benchmark possible, but they also inherit pretraining biases that may affect crops, captions, VQA answers, and embeddings. Automated metrics cannot capture every subjective harm, microaggression, or culturally situated stereotype.

The benchmark also does not include human annotation by people with disabilities, because of scale. It does not cover all functional disability groups, assistive technologies, or contexts. The authors point future work toward PWD-centered evaluation. INCLUDE-BENCH is a serious diagnostic, not a substitute for affected-community judgment.

Source Discipline

Primary sources were the arXiv abstract, API record, PDF, and experimental HTML. This page paraphrases the paper and does not reproduce figures, tables, sample images, or long passages. Benchmark size, model count, prompt subsets, metrics, findings, and limits come from those records.

Sources

• Sophia Lichtenberg, Albert Gatt, and Judith Masthoff, Beyond wheelchairs and blindfolds: Investigating disability stereotypes in T2I models with INCLUDE-BENCH, arXiv:2607.08515 [cs.CV], submitted July 9, 2026.
• arXiv API record for arXiv:2607.08515, checked for title, authors, category, submission date, and version metadata.
• arXiv PDF for arXiv:2607.08515, checked for page count, affiliations, benchmark construction, model list, results, and limitations.
• arXiv experimental HTML for INCLUDE-BENCH, checked for section structure, table text, model references, and limitation wording.

Return to Blog