Sensitive content models separate risk from noise at 80%+ accuracy

This post details the development and capabilities of sensitive content models that enhance Glean's unstructured data governance. These models combine traditional infotype classifiers with AI models trained on enterprise context (document content, activities, permissions) to achieve over 80% accuracy in identifying sensitive data and distinguishing benign use cases from true exposure. The models leverage semantic relationships between document titles, content summaries, and identified sensitive patterns, as well as contextual clues like 'example' or 'sample'. They also analyze document linkage, folder structures, ownership, and access patterns, integrating with the enterprise graph to understand the document's role and access patterns within the organization. The system uses in-product feedback (thumbs up/down) for continuous learning.