New methods of art authentication rarely receive a warm welcome from the established art world. Connoisseurs have spent decades training their eyes, studying artists’ habits and learning how workshops operated. A computer trained to examine paintings can seem like an intruder at the table.
Professor Nils Büttner had serious reservations. One of the world’s leading authorities on Anthony van Dyck, he published a pointed critique of AI-based authentication and questioned whether the technology could provide the kind of evidence art historians require.
Art Recognition could have treated the criticism as hostility. Instead, the team opened its doors.
Over a series of conversations, Art Recognition explained how its models were trained, how its art historians selected and checked reference images, and how the results were interpreted. Büttner brought the scrutiny of a catalogue raisonné scholar: Which paintings belonged in the dataset? Which workshop pictures might contaminate it? How reliable were the accepted attributions on which the model depended?
The discussion soon found a perfect test case.
The painting was a portrait of the Marqués de Leganés, long associated with Van Dyck. Its precise status remained uncertain. Was this a portrait painted by Van Dyck himself, or a version produced in his workshop?
Büttner and Art Recognition agreed to investigate the painting independently. First, they revisited the material used to train the Van Dyck model, checking it against the authoritative Barnes catalogue raisonné and refining the dataset. Art Recognition then analysed the portrait with its artist-specific model. Büttner carried out his own examination using art-historical, stylistic and technical evidence.
Both investigations arrived at the same answer.
Art Recognition’s model returned a 79 per cent probability against attribution to Van Dyck himself. Büttner’s examination also identified the painting as a workshop variant. Two very different methods, applied separately, had reached the same conclusion.
The results led to a jointly authored, peer-reviewed article in the Kunstgeschichte E-Journal in 2024. Büttner had entered the discussion as one of AI authentication’s most prominent critics. He emerged as a collaborator whose questions had helped strengthen the model and its underlying research.
That journey matters as much as the verdict on the painting. Art historians are right to question new authentication tools, especially when an attribution may determine a work’s place in a catalogue, a museum or the market. Trust has to be earned through access to the method, careful datasets and results that other specialists can examine.
The Leganés portrait offered a practical demonstration. AI supplied a measurable analysis of the image. Connoisseurship supplied historical knowledge, technical judgment and an expert eye trained over decades. Used together, they produced a fuller and more accountable assessment than either could have offered alone.
Sometimes the most valuable critic is the one willing to stay for the experiment.