MAI models use common crawl despite 'clean data' promises

microsoft's technical paper for maia-1 reveals that the models were trained on common crawl and other publicly available, unlicensed web data. this finding contradicts microsoft's explicit marketing promises that the models were built using only enterprise-grade, clean, and commercially licensed datasets. while microsoft claims to respect robots.txt and standard web protocols, it effectively shifts the burden of content protection to site owners rather than securing explicit licensing. this highlights the ongoing discrepancy in the industry between marketing 'clean' ai products and the legal reality of relying on fair use for training data.