MCN 2026

In recent years there has been much productive discussion both at MCN and within institutions on AI-generated visual descriptions for accessibility: how to generate descriptions, how good they need to be, and how institutions should oversee their use. Some museums have already successfully released AI-generated descriptions as accessibility text. 

The Met has been treating AI-generated descriptions as something different: not (yet) as a final user-facing product, but as raw search material. With descriptions of what artworks depict, embedded into vector space and indexed alongside traditional metadata, one can build a discovery experience that works not only for experts, but also for the roughly two-thirds of users who arrive at our collection without a specific artwork in mind. Users searching for "year of the horse," "mythical beasts," or "eating food together" can find relevant works for the first time. Users searching in Chinese, Spanish, or any of dozens of languages can now find the same works as users searching in English. Users without art-historical vocabulary find what they are looking for without needing to know an artist's name or an artwork’s title. And yes, the search still works for our experts and researchers who search via fields like accession number.

We will share what we learned shipping this to a production system of The Met’s 550,000 objects used by millions of visitors a year. We will cover the technical pipeline (vision-language models, embeddings, hybrid lexical-and-semantic search), the evaluation problem (with different concerns when descriptions are an input to search rather than user-facing text), the UX work (in many ways more difficult than the engineering), and what our post-launch analytics and user feedback have actually shown. We will be candid about what worked, what surprised us, what still fails, and what we would do differently. We will also share what transfers to other institutions and what is specific to working at this scale.