Bridging the Data Provenance Gap Across Text, Speech and Video

Web-scraped and social media sources now dominate AI training data: Since 2019, multimodal datasets have overwhelmingly relied on web-crawled, synthetic, and social media platforms like YouTube, replacing all other data sources as the primary training material for AI systems.

Licensing restrictions are more pervasive than they appear: While fewer than 33% of datasets carry restrictive licenses directly, over 80% of the actual source content in widely-used text, speech, and video datasets contain non-commercial restrictions, creating a significant gap between dataset-level and content-level licensing compliance.

Geographic and linguistic diversity has stalled despite appearances of progress: Despite an increase in the absolute number of languages (608) and countries (67) represented in public datasets, the audit found that measures of relative representation have failed to meaningfully improve since 2013, suggesting that Western-centric data dominance persists.

The audit reveals systemic transparency gaps across AI modalities: This first-of-its-kind longitudinal analysis of nearly 4,000 datasets across text, speech, and video (1990-2024) demonstrates that the AI industry lacks comprehensive visibility into data provenance, sourcing practices, and their implications for responsible AI development.