Can an AI authenticate a painting better than an expert's eye, and who wins when they disagree?

Not on its own, so far. In two disputed cases, an AI verdict has clashed with tree-ring evidence and most scholars, or with another AI, so no single method wins. Three kinds of evidence compete: the expert's eye, science, and now AI.

The expert's eye is called connoisseurship. Connoisseurs evaluate works of art on the basis of their experience of the style and technique of artists. In the art trade it remains a crucial skill for attributing works by style and technique where documentary evidence of provenance is lacking. The Rembrandt Research Project was considered the authority on Rembrandt's paintings, and its research reduced the number of signed Rembrandt self-portraits around the world by half. Yet the project also started a debate about the feasibility of conclusive attribution, particularly for painters associated with one or more workshops.

Science and the eye can disagree. At a 1992 conference on the Getty kouros, most art historians and archaeologists denounced the statue while the scientists present believed it authentic.

AI is the newest voice. Art Recognition, a Swiss company, specialises in applying artificial intelligence to art authentication. In November 2024, Germann Auction became the first auction house to conduct a sale of artwork authenticated entirely by artificial intelligence. But in September 2021, an AI analysis announced a 91% probability that the National Gallery's Samson and Delilah was not by Rubens, while a tree-ring examination confirmed the right period and a majority of scholars accept the attribution. On another painting, the De Brécy Tondo, two AI systems gave opposite answers.