Beyond Wheelchairs and Blindfolds Investigating Disability Stereotypes in T2I Models with INCLUDE-BENCH

Jul 16, 2026

Text-to-image (T2I) models can reproduce social biases, yet disability stereotyping has been insufficiently evaluated with sociologically grounded definitions. The paper introduces INCLUDE-BENCH, a large-scale benchmark for measuring disability-related bias in T2I systems. INCLUDE-BENCH contains 119K generated images from prompt-designed bias dimensions in static and dynamic contexts. Experiments across 15 open and 2 closed models show wheelchair-centric depictions from mobility-impaired/default prompts, reduced diversity under disability-conditioned prompts, and stronger disability–text alignment with stereotyping, quantified via an SCM-based score.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages13

Reading Time33 min

Abstract

Introduction

Disability bias in T2I models

Stereotypes and representativeness framework

INCLUDE-BENCH contributions

Beyond wheelchairs and blindfolds: Investigating disability stereotypes in T2I models with INCLUDE-BENCH Sophia Lichtenberg, Albert Gatt, Judith Masthoff Utrecht University [slichtenberg@uu.nl](slichtenberg@uu.nl) arXiv :2607 .085 15v 1 [ cs .CV] 9 Jul 2026 Abstract Text-to-image (T2I) models have been shown to exhibit social biases. Prior work has mainly focused on gender, skin tone, and cultural representation within restricted occupational associations, and emerging benchmarks increasingly incorporate these dimensions. However, disability remains systematically underexplored. Current evaluation practices often fail to align with sociologically grounded definitions of stereotyping, limiting principled assessment of representational harms toward people with disabilities (PWD). To address this, we introduce INCLUsive Disability Evaluation (INCLUDE-BENCH), the first largescale benchmark for evaluating disability-related bias in T2I models. INCLUDE-BENCH comprises 119K generated images based on prompt design, across multiple bias dimensions and both static and dynamic contexts. We evaluate 15 open-source and 2 closed models. Our key findings reveal that: (1) mobility-impaired and default disability prompts predominantly yield wheelchair depictions across all models; (2) disability-conditioned generations consistently exhibit less diversity and (3) stereotypical portrayals demonstrate stronger disability–text alignment and (4) introduce Stereotype Content Model (SCM) Score, demonstrating that T2I models reflect real-world stereotypical associations. 1. Introduction Text-to-image (T2I) models, which synthesize images from textual descriptions, have rapidly advanced in both capability and popularity, and are increasingly used in digital content creation and other creative domains. Despite these advances, T2I models often reproduce societal biases present in their training data. While gender, age, race, and cultural biases are relatively well studied [100, 27], other marginalized groups and intersectional identities remain underexplored. T2I models also raise concerns around toxicity, fairness, and authenticity, with representational harms that can stereotype, erase, or demean social groups, limiting individuals’ control over their depiction and contributing to negative cognitive and emotional outcomes [4, 11, 12, 57, 87, 101] . Disability representation has received limited attention, despite affecting a substantial portion of the population: an estimated 16% globally [106], and approximately 28.7% of U.S. adults, with cognitive (13.8%) and mobility (12.2%) impairments most common [19] . Prevalence also varies by gender, with 20% of women and 12% of men living with a disability [96] . Yet representation remains minimal: a 2024 study found that only 2% of respondents with disabilities felt represented in the media, indicating that accessible experiences, accurate depiction, and authentic narratives remain largely unmet [94] . Existing studies show that T2I models frequently produce stereotypical and narrow portrayals of persons with disability (PWD), often depicting them as older, sad and reliant on old-fashioned aids (e.g. manually operated wheelchairs) . Communitycentered analyses further reveal inaccurate, unsafe, and homogeneous depictions, persistent reliance on wheelchair tropes, erasure of multifaceted identities, and failures to represent certain type of disablity [93, 6, 66] . In sociology, stereotypes are defined as generalized beliefs about social groups,“mental representations of real differences between groups”, that simplify information processing [14] . The representativeness framework [14] states that stereotypes arise from selective recall of the most distinctive, diagnostic features, which maximize between-group differences and exhibit low within-group variation [51], often compensating for limited information. From this perspective, T2I models over-weight highly diagnostic visual features,

FAQs

What gap does INCLUDE-BENCH address in prior disability bias evaluation for T2I models?

Prior work has not adequately aligned evaluation with sociologically grounded definitions of stereotyping, limiting principled assessment of representational harms toward people with disabilities.

How is INCLUDE-BENCH constructed and what scale does it cover?

INCLUDE-BENCH is built by curating multiple disability function-group prompts across subsets and generating 20 images per prompt across 17 state-of-the-art T2I models, totaling 119,680 images.

What key behaviors do the results reveal about disability prompts and stereotyping?

Mobility-impaired and default disability prompts tend to produce wheelchair depictions across models; disability-conditioned generations show reduced diversity; and stereotypical portrayals exhibit stronger disability–text alignment, captured using an SCM-based score.

Jul 16, 2026

Upload by Valentina