Artificial intelligence casts doubt on Jan van Eyck’s authorship:
two "Saint Francis" paintings are in question.

In February 2026, the Swiss company Art Recognition announced that two works, classified for centuries as works by the Flemish artist Jan van Eyck, were not by the master. Both paintings are housed in major museums in Philadelphia and Turin. The academic community responded not with agreement, but with questions.

Two almost identical canvases

These are " Saint Francis of Assisi Receiving the Stigmata " — two nearly identical depictions of the same scene, dating from approximately 1428 to 1432. One is housed in the Philadelphia Museum of Art, the other in the Savoy Gallery of the Royal Museums of Turin. Both paintings are unsigned and are among the few works traditionally attributed to van Eyck.

The Turin and Philadelphia versions are similar but not identical. The Turin version is smaller. The background details and the modeling of the figures differ — and these differences have long fueled debate: whether both works are original, one a master copy, or both painted without van Eyck’s direct involvement. Art historians have been debating this issue long before any algorithms were invented.

What the computer showed

Art Recognition is a Swiss company founded in 2019 in Zurich. Its founder and director is Dr. Karina Popovic. The company specializes in authenticating artworks for the art market and collaborates on research projects with Tilburg University in the Netherlands.

The results of the analysis were definitive: the Philadelphia version received a " 91% negative " rating for van Eyck’s authorship, while the Turin version received an " 86% negative " rating. In other words, the algorithm did not detect any patterns in the paintings that are characteristic of the master’s brushwork.

Popovich stated that museums "would not be happy" with such findings. Neither the Philadelphia Museum of Art nor the Royal Museums of Turin responded to journalists’ inquiries.

Analysis technique

The Art Recognition algorithm is built on two components: a convolutional neural network (CNN) and a vision transformer with shifted windows . The CNN identifies local textural patterns — paint texture and brushstroke patterns. The transformer analyzes larger spatial structures in the image.

The company’s specialists began with high-quality digital scans of both paintings. The program divided each canvas into hundreds of thousands of small fragments. The algorithm analyzed each fragment based on several parameters: brushstroke curvature, pressure distribution, texture frequencies, and edge definition. The resulting data was compared with a database of verified works by van Eyck.

The sheer size of this database was the first stumbling block. Around twenty-five paintings is a paltry number for reliable deep learning.

Jan van Eyck: The Artist and His Time

Jan van Eyck was born around 1390 and died in 1441 in Bruges. He served as court painter and diplomat to Philip III the Good, Duke of Burgundy. Van Eyck did not invent oil painting; he brought the technique to an unprecedented level of precision and luminosity.

Among his most famous works are the " Ghent Altarpiece ," created in collaboration with his brother Hubert, and the portrait " The Arnolfini Couple ," now housed in London’s National Gallery. His works were distinguished by the microscopic detail of surfaces — glass, fabric, leather — and the psychological precision of his portraits. Van Eyck’s work had a direct influence on Italian painters, including Antonello da Messina.

The body of works attributed to him is small — around twenty-five paintings. For a master of his stature and stature, this is exceptionally small. Some of the studio’s works exist under other names, while others were lost during wars and displacement.

The workshop as a norm of the era

In the 15th century, the concept of a single, authorial work was much less rigid than it is today. Most major painters ran workshops in which apprentices performed a significant portion of the work. Backgrounds, clothing, architectural details, minor figures — all of this could be painted by an apprentice under the supervision of a master.

The scene of the "Stigmatization of Saint Francis" was particularly popular among Europe’s Catholic aristocracy. Multiple versions of a single composition were common practice during the era. The original was often reproduced by assistants; in some cases, more than two versions were created. In this context, the question of whether van Eyck painted it takes on a different aspect: perhaps he developed the concept and executed the key parts, leaving the rest to his workshop.

The specifics of the two paintings

Beyond the general question of the workshop, both versions have specific physical characteristics that significantly complicate computer analysis.

The Philadelphia painting is painted on parchment glued to a wooden support. This is an atypical medium: the surface texture is radically different from van Eyck’s traditional oil paintings on limestone ground. The algorithm, trained to recognize van Eyck’s brushstrokes on a familiar support, in this case analyzes a fundamentally different physical object.

Both works have been damaged and restored numerous times over six centuries. A significant portion of the paint layer analyzed by the program may not be part of the original paint layers, but rather later overpaintings. Maximilian Martens of Ghent University posed a direct question: "Does Art Recognition take the paintings’ state of conservation into account?"

Restoration interventions alter the texture and color characteristics of a surface. If the algorithm is trained on clean scans of well-preserved works, and the painting under analysis has undergone several restorations, a negative result may be due not to van Eyck’s lack of authorship, but to paint degradation and the accumulation of foreign paint.

Voice of the Scientist: Maximilian Martens

Martens, one of the recognized experts on van Eyck, publicly rejected the findings of Art Recognition. His key objection concerns the nature of the master’s painting itself.

"Even when studying van Eyck’s paintings under a microscope, his brushstrokes are barely discernible. This is one of the most characteristic features of his style," he wrote in an email. In other words, the task of "recognizing the artist’s brushstrokes" is particularly difficult with van Eyck: there are almost no brushstrokes as such. The paint is applied in thin, translucent layers, and the line between the master’s work and that of his apprentice is virtually invisible on the surface.

Twenty-five paintings are insufficient to train a model, Martens believes, "even if every square centimeter of the surface is taken into account." In addition to the small sample size, Martens pointed to the lack of peer-reviewed scientific publications describing the company’s methodology. Without an open methodology, third-party researchers cannot reproduce the analysis, validate the training set, or assess the model’s systematic biases.

No recognized van Eyck scholars were involved in the study. Martens put it bluntly: "These algorithms must be trained by art historians and restorers who have dedicated decades to studying van Eyck."

Till-Holger Borchert: AI data fits into the existing debate

A different position was taken by Till-Holger Borchert, one of the leading van Eyck experts and director of the Sürmondt-Ludwig Museum in Aachen. According to him, Art Recognition’s findings are consistent with those of researchers who had previously believed that both versions of "Stigmatization" were masterpieces — painted in the artist’s studio, but not necessarily by his hand.

Borchert didn’t see the results as a sensation. For him, they’re an additional argument in a long-running and unresolved debate, not a refutation of anything established.

Noah Charney: The Lost Original

Art historian Noah Charney discussed the preliminary results of the Philadelphia painting’s analysis on his podcast before the final results were published. Charney described Art Recognition’s previous analyses as "remarkably accurate" and noted that the negative results for both paintings were so unexpected that they required further testing for confirmation.

Charney expected something different: the Turin version would be confirmed as authentic, while the Philadelphia version would be a copy — a workshop work or a later imitation. Having received negative results for both, he proposed an alternative explanation: "The negative results indicate that both works are masterpieces. This may indicate the existence of a lost original, closer to the hand of van Eyck himself."

The hypothesis of a lost original is not an exotic idea. The history of 15th-century Netherlandish painting is replete with cases where only copies survived, while the original did not survive wars, fires, or simple neglect. Many works from the early Netherlandish period have come down to us only as later reproductions.

The Rubens affair and the question of reputation

Art Recognition has attracted public attention before. In 2021, the company declared that the painting "Samson and Delilah," held at London’s National Gallery and attributed to Rubens, was a forgery. Leading Rubens experts rejected this conclusion.

Niels Büttner, chairman of the Centrum Rubenianum in Antwerp and the head of the Rubens catalogue raisonné, dismissed doubts about the painting’s authenticity as "conspiracy theories." Martens is convinced that this very story has undermined Art Recognition’s credibility in academic circles. "I fear that the media storm they are currently unleashing will further damage their reputation," he wrote. Popović declined to comment on the matter.

Over the years, the company has also worked on other attributions, analyzing works attributed to Monet, Renoir, Raphael, and Caravaggio. However, it was the controversies surrounding Rubens and now van Eyck — with their widespread media attention — that have raised questions about the academic community’s attitude toward its methods.

What remains outside the algorithm

Martens and other experts point out that art attribution is a discipline that requires a thorough knowledge of style, an understanding of the object’s physical history, familiarity with archival sources, information about the client and patron, and the ability to work with pigment chemical analysis and X-ray imaging. An algorithm that works solely with a digital image of the surface obtains only a minimal portion of this information.

It doesn’t take into account restoration interventions. It doesn’t see X-rays. It doesn’t read van Eyck’s correspondence with the Duke of Burgundy, and it knows nothing about the history of a particular painting. "AI analyzes texture, not history," as one commentator on this discussion put it succinctly.

The very lack of consensus among historians regarding "Stigmatization" suggests there is no simple answer. Martens has explicitly stated that there is no "consensus" regarding either version. This fact calls into question the very premise — that an algorithm can resolve what experts working with primary sources have been unable to do for decades.

Charney and Martens have different opinions on the Art Recognition methodology. But both agree on one thing: a negative algorithm result doesn’t resolve the question of authorship — it returns it to the desks of historians, restorers, and physical research methods that work with the paintings themselves, not their digital copies.