The fantasy is simple: point an algorithm at a collection and let it assemble a show—connections no one’s noticed, themes no one’s named. In practice, the “AI curator” feels less like a ghost replacement and more like a new intern with superpowers: fast, tireless, occasionally brilliant, and usually in need of a firm, human hand. When Duke’s Nasher Museum asked a model to propose an exhibition, they installed the results with the process on view; the value wasn’t abdication but perspective. The system surfaced overlooked works and odd pairings, while also showing how brittle machine criteria become under cultural pressure. Elsewhere, institutions are treating AI less as a decider and more as a lens. MoMA’s collaborations with Refik Anadol, for instance, let a model “dream” the archive—refracting a century of art into moving weather rather than picking objects like a checklist. At infrastructure scale, big collections are wiring themselves for machine-speed wandering: interfaces that map latent patterns across centuries, tools that pair artworks by visual rhyme alone. Useful? Absolutely. They widen the searchlight and put fresh proposals on the table at 2 a.m.