Arts — Noozify Original — August 14, 2026

The Eye, the Algorithm, and the Fight Over What Counts as a Van Eyck

By Adam Canissario

For roughly a century, two paintings of Saint Francis of Assisi Receiving the Stigmata—one in Turin, the other in Philadelphia—have carried Jan van Eyck's name. That changed this year when Swiss company Art Recognition analyzed both with an AI model and concluded that neither was by Van Eyck, assigning the results as much as 91 percent certainty. Maximiliaan Martens of Ghent University, a leading Van Eyck authority, was unconvinced. Van Eyck's technique is so subtle that even microscopes struggle to reveal his brushwork. What, then, had the algorithm actually learned? More fundamentally, there is no agreed scholarly baseline for these panels from which the system could reliably learn.

The stakes are high because the corpus is tiny. About twenty surviving paintings are confidently given to Van Eyck, all dated between 1432 and 1439, plus the Ghent Altarpiece; the nine portraits the National Gallery in London will gather from November 21st amount to nearly half his oeuvre. None comes to auction. The two Saint Francis panels belong to the Philadelphia Museum of Art and Turin's Galleria Sabauda, so a new label sells nothing, yet it still moves the market. A demoted panel loses loans, insurance value and its place in the literature; a promoted one becomes an asset of another order, which is why owners of plausible followers now want an algorithm's opinion.

Traditional connoisseurship began as an eye and a hunch. Giovanni Morelli, a 19th-century Italian with medical training, argued that a painter's habits show most in details he never thinks about, the ears, hands, and fingernails, and that those tics separate original from copy. Provenance adds the paper trail through sales and inventories. The laboratory then tests both. Tree-ring dating can fix when a Netherlandish painter's oak boards were felled, since dendrochronology works on oak but not on poplar, so a board cut after Van Eyck's death in 1441 ends the argument. Infrared reflectography relies on the fact that carbon black absorbs infrared radiation, so the underdrawing shows through the paint, along with every change of plan. The benchmark is Brussels, where KIK-IRPA, the Royal Institute for Cultural Heritage, has led the restoration of the Ghent Altarpiece since October 2012 and found around 70 percent of the outer panels hidden under 16th-century overpainting, including the eerily human face of the Lamb, and produced the documented baseline the Saint Francis panels lack.

The Saint Francis disagreement exposes the central problem. AI can produce an impressive probability, but that number is only as meaningful as the evidence and methodology behind it. Scholars can explain their reasoning; proprietary algorithms often cannot. We have encountered this problem with other opaque systems, but the stakes are different here. The machine is not merely assisting researchers. It is challenging judgments built through centuries of artistic scholarship.

Art Recognition faced a similar dispute last year over Rubens's The Bath of Diana. Its model identified the artist's hand in a painting long regarded as a copy. Nils Büttner, chairman of the Centrum Rubenianum and curator of Rubens's catalogue raisonné, rejected the conclusion. According to him, the painting's condition ruled out an authentic Rubens. He also pointed to a flaw in the training material: the dataset did not include the composition scholars recognize as the relevant studio version. The system may have been highly confident, but confidence cannot compensate for incomplete evidence. A precise percentage can look scientific while still being fundamentally wrong; precision has never guaranteed accuracy.

Art Recognition's earlier verdicts deserve the same scrutiny. In 2021 it analyzed a digital reproduction of the National Gallery's Samson and Delilah and reported a 91.78 percent probability that Rubens did not paint it, after training on 148 uncontested works. The museum, which had published a full technical study of the panel in 1983, said it would wait for the research to be published in full. Two years later a facial-recognition system from the University of Bradford declared the de Brécy Tondo a Raphael because its Madonna matched the Sistine Madonna at 97 percent. One Raphael specialist told The Art Newspaper the picture was "clearly a copy from the 19th century", which is precisely why the faces match. Art Recognition's own model put the same tondo at 85 percent not Raphael. Two machines, opposite verdicts, no way for an outsider to adjudicate.

Human experts, however, hardly have an unblemished record. The Rembrandt Research Project repeatedly revised its opinions about individual works, illustrating how attribution can change as scholarship evolves. The Van Eyck and Rubens disputes therefore should not be reduced to machines versus humans. The unresolved issue is how AI-generated conclusions should be treated as evidence. Are they laboratory findings, expert opinions, or sophisticated guesses? The art market has yet to settle that question, even though the financial consequences can be enormous.

A fair standard for machine evidence is easy to describe; nobody has adopted it. Publish the training set, so specialists can check whether the relevant studio versions were in it. Report performance on known cases, including workshop pieces and period copies, not just autograph works against crude fakes. Say whether the system saw the painting or a photograph, and at what resolution. Then give the result the status dendrochronology has earned, a constraint rather than a verdict; a felling date can exclude an attribution but never confirm one. Held to that standard, a 91 percent figure becomes one entry in a dossier beside the underdrawing and the oak. Held to none, it is a press release.

Meanwhile, criminals are using generative AI for a more straightforward purpose: fraud. Traditional counterfeiters had to create a convincing object. Now they can fabricate the paperwork instead. AI can generate appraisals, invoices and provenance documents that appear credible enough to survive an initial review. One insurance investigator examining a claim involving a major collection found dozens of seemingly legitimate certificates before metadata exposed the deception. The collection itself appeared to be fictional. There was no forged canvas to inspect—only an invented history surrounding nonexistent art.

That development reveals the larger vulnerability. The art trade depends heavily on documentation and trust, and AI is now attacking both. On one side, it can manufacture convincing evidence for objects that never existed. On the other, it can generate authoritative-sounding judgments about disputed works. Companies developing provenance systems and knowledge graphs may help strengthen the paper trail. But the fundamental question remains unchanged: when evidence is incomplete or ambiguous, who gets the final say? For now, experts still hold that position—but their authority is increasingly contested. AI has not replaced connoisseurship. Instead, it has exposed just how much authentication has always depended on trust.