Data Representativeness in Accessibility Datasets: A Meta-Analysis

Accessibility datasets show significant representation gaps: While datasets sourced from people with disabilities and older adults demonstrate diverse age representation, they have notable gaps in gender and race & ethnicity representation, which could perpetuate bias in AI systems designed for inclusive applications.

Demographic labeling lacks consistency and transparency: The paper identifies a critical problem where demographic variables like gender and race & ethnicity are classified inconsistently across the 190 datasets reviewed, and the source of these demographic labels is often unknown, making it difficult to assess data quality and reliability.

Current accessibility datasets may not serve their intended purpose: Despite being specifically designed to improve AI fairness for marginalized groups, these datasets themselves underrepresent certain demographic groups, potentially limiting their effectiveness in creating truly inclusive AI systems.

Demographic data collection needs standardization: The research highlights that the sensitive nature of collecting demographic information from disabled individuals requires more transparent, consistent, and documented approaches to ensure both ethical data practices and usable datasets for fairness research.

Representation gaps risk compounding existing inequities: Without addressing underrepresentation in accessibility datasets, AI systems trained on this data may continue to produce unfair outcomes for disabled individuals from underrepresented racial, ethnic, and gender groups—defeating the goal of building more inclusive technology.