Art history is full of doubles. Copies, replicas, studio versions, later repetitions. Most of the time, they coexist peacefully. But every so often, two nearly identical works surface that force the art world into an uncomfortable binary. Only one can be right. And sometimes, neither institution nor market seems eager to say why.

Kees van Dongen was a central figure among the Fauves, celebrated for his saturated colors, stylized figures, and unapologetically modern approach to the female nude. His early work absorbed influences from the Hague School and Symbolism before evolving into a roughened, expressive painterly language that balanced elegance with provocation. It is a style that can look deceptively simple, and for that very reason, dangerously easy to imitate.

Over the past two decades, two almost indistinguishable versions of a painting titled Nu en Buste have appeared on the art market. For years, they circulated quietly, each carrying institutional legitimacy, until their coexistence became impossible to ignore.

Version A surfaced publicly in 2008, when it was sold at auction accompanied by a certificate of authenticity issued by the Wildenstein Institute. At the time, this was as close to a seal of approval as the market could hope for. Yet when the same painting reappeared for auction in Paris in December 2019, the sale was abruptly canceled. By then, responsibility for van Dongen’s catalogue raisonné had passed to the Wildenstein Plattner Institute, which rejected the earlier appraisal and refused to include the painting in the catalogue. No explanation was offered.

Just months later, in May 2020, Version B was auctioned in Paris. This time, the painting carried a certificate from the WPI itself. In a twist that only deepened the confusion, the institute indicated that it might reconsider its verdict on Version A, provided the owner could supply more complete provenance documentation. That proved impossible. The owner could trace the painting only as far back as the 2008 auction and a mention in a 1989 catalogue. The auction house responsible for the earlier sale was unable to provide additional records. The result was a familiar art-world impasse: one painting approved, one rejected, and no transparent rationale separating the two.

Faced with this uncertainty, the owner of Version A turned to Art Recognition, seeking a form of analysis that operated outside institutional opacity and market politics.

The first step was to teach the AI how van Dongen actually paints. Using high-resolution images of confirmed works, we assembled a training dataset drawn from major exhibition catalogues, including those of the Musée National d’Art Moderne in Paris (1967), the Museum Boymans–van Beuningen in Amsterdam (1968), and the Fondation Pierre Gianadda in Martigny (2002). To prevent the system from merely learning “Fauvism” in general, we also included works by stylistically adjacent contemporaries, notably Otto Müller and Albert Marquet, whose visual languages overlap just enough to confuse a less rigorous model.

Once trained, the AI analyzed both versions of Nu en Buste. The results were unambiguous. Version A was classified as a forgery with a 73 percent probability. Version B, by contrast, was classified as authentic with an 81 percent probability. This was not a subtle distinction, nor one driven by a single feature.

To understand where the differences lay, we conducted a patch-by-patch analysis, dividing each painting into discrete visual zones and assessing each independently. In Version A, the least authentic areas clustered around precisely those passages where van Dongen’s hand matters most. The breasts registered as 91 percent and 87 percent not authentic. The hair showed a 75 percent probability of inauthenticity. The upper face scored 85 percent not authentic. These are not marginal details. They are the structural core of the image, where gesture, rhythm, and painterly confidence reveal themselves most clearly.

The market outcome mirrored the analysis. Version B proceeded to sale. Version A did not. What remains is a remarkable imitation, close enough to pass through institutional hands once, but not close enough to withstand sustained scrutiny from multiple angles.

This case illustrates something essential about authentication in the modern art market. Certificates can change. Institutions can reverse themselves without explanation. Provenance can evaporate. What AI offers is not authority, but consistency. In the case of Nu en Buste, it provided a clear, replicable distinction between two nearly identical paintings, when human systems proved unable, or unwilling, to do so.

< Back