Past Meets Future

Libraries, archives, and museums are widely trusted as stewards of authoritative records. At the same time, the data underlying these collections is heterogeneous, shaped by different institutional practices, and often tied closely to the historical and scientific contexts in which it was created. As institutions begin to apply AI methods to support discovery and interpretation, a central challenge emerges: how do we maintain trust in authoritative records while inviting users into the uncertainty and iteration of AI-driven processes?

This presentation draws on lessons learned from AI pilot projects across the Smithsonian's science and history museums and research centers, with a special focus on the recent Revolution Crossroads project. Revolution Crossroads is a multi-institutional, public-facing initiative bringing together collections data from several large, national-level cultural heritage organizations. Working across datasets from the American Revolutionary War era (1770–1810) that differ in format, structure, standards, and historical context, the project is using AI to explore connections at scale.

Across these pilots, unsupervised AI methods have sat comfortably alongside institutional expertise because they organize rather than assert. I will share examples of how multimodal embedding models have assisted in collections discovery, evaluation, and even accessible alt-text image description. Throughout, the goal is not to replace curatorial judgment but to make it more visible — revealing where models are confident, where they are guessing, and where expert intervention matters most. Trust, in this framing, comes not from hiding uncertainty but from designing systems that keep experts and users in the loop.