SoUnD Framework: Analyzing (So)cial Representation in (Un)structured (D)ata

Unstructured data poses hidden representation risks: Foundation models are trained on massive unstructured datasets like web text and images, making it difficult to systematically understand how different groups of people are represented—a critical gap for responsible AI development.

The SoUnD Framework provides a structured approach: The paper introduces a practical framework specifically designed to help organizations analyze human representation in unstructured data and identify potential downstream harms before deployment.

Real-world datasets show representation challenges: Testing on Common Crawl (C4) and LAION-400M revealed that systematic analysis of these widely-used datasets is feasible, suggesting organizations can implement similar audits of their own training data.

Documentation and decision-making require evidence: The framework supports more informed choices about dataset use and development by providing concrete methods to document how people are represented, moving beyond assumptions to evidence-based decisions.

Actionable steps exist for responsible data practices: The paper proposes specific action steps for dataset use, development, and documentation, indicating that responsible AI practices around data representation are implementable rather than purely theoretical.