Fallacies of Data Transparency: Rethinking Nutrition Facts for AI

Policy mandating data transparency has recently emerged as an attempt to address concerns around generative AI such as data quality, privacy, and copyright. However, there are significant gaps between the aspirations of data transparency and what specific measures can actually achieve. In this work, we identify the challenges in achieving the intended goals of providing nutrition facts for AI without the necessary disclosures, enforcement mechanisms, and pathways to change. We provide an institutional perspective on three fallacies in calls for data disclosures. First, there is a \textit{disclosure gap} between the stated goals of data transparency and the disclosures that would be required to achieve these goals. Second, there is a \textit{remediation gap} between required disclosures and the enforcement mechanisms that ensure compliance. Third, there is an \textit{outcome gap} between the appearance of compliance and the intended behavior change in AI system developers. Our analysis identifies shortcomings in existing data transparency policy that need to be addressed in order to ensure genuine accountability.