Search

Representational harm in materials dataset construction remains a critically overlooked failure mode in artificial intelligence for materials science, where systematic patterns of underrepresentation and misrepresentation silently shape which materials are studied, discovered, and deployed while rendering entire classes of materials, synthesis pathways, and knowledge traditions invisible to AI systems. Representational harm is defined here as the systematic underrepresentation, misrepresentation, or exclusion of certain material classes, synthesis methods, research traditions, or communities within materials datasets, resulting in biased AI models that perpetuate inequitable discovery outcomes and reinforce existing power structures in the field. This article articulates five distinct types of representational harm—chemical, structural, synthetic, geographic, and community—along with the four primary mechanisms through which dataset construction choices actively produce these harms, including historical priority, funding asymmetry, measurement accessibility, and publication bias. It further presents a typology of four specific harm failure modes that emerge in materials AI pipelines: invisible classes, distorted property distributions, representational feedback loops, and knowledge colonization. Finally, the paper offers practical detection principles based on diversity, geographic, citation, and community audits as well as five mitigation principles centered on intentional dataset design, data enrichment, weighted representation, inclusion of multiple knowledge systems, and ongoing harm auditing, thereby providing a comprehensive framework for transforming materials dataset construction into a more equitable and epistemically responsible practice.