When it comes to A.I. authentication, the importance of good data means cases are rarely cut and dry.
A version of Rubens's composition The Bath of Diana from a private collection in France. Image courtesy Art Recognition.
A new artificial intelligence analysis has reignited debate over the authenticity of The Bath of Diana (ca. 1635), a painting long believed to be a copy of a lost Rubens work. A Swiss authentication firm claims that some areas of the painting bear the Flemish master’s hand—though a leading Rubens scholar strongly disputes the attribution. The case underscores both the potential and the pitfalls of A.I.-driven authentication in the art world.
The Swiss tech start-up Art Recognition presented its findings from a recent analysis of the work at the Art Business Conference in Maastricht last week. “We conducted an A.I. authenticity analysis on individual patches, and the results were mixed: while most patches were identified as authentic, some were not,” Art Recognition’s CEO and co-founder Carina Popovici explained by email. “This means the painting was not created entirely by Rubens.”
Of 29 patches of the painting tested by Art Recognition’s A.I. model, ten were authentic with a probability higher than 80 percent, eight were authentic with a lower probability between 60 to 80 percent, seven were inconclusive, and four were determined to be inauthentic. One of these latter patches included the central figure of Diana. The differing probabilities could be explained by the fact that Rubens is known to have worked with a workshop of assistants.
The analysis was a paid assignment commissioned by the French collector who owns the painting. Art Recognition has completed over 500 analyses like these for private clients in over 20 countries, according to the firm.
The real test for A.I. authentication is whether it can have an influence on the market and last year the Swiss company backed the first A.I.-authenticated work to sell at auction.
The judgement about the French The Bath of Diana has already been called into question by Nils Büttner, as first reported in the Art Newspaper. An outspoken supporter of A.I. authentication tools, Büttner is one of the leading authorities on Rubens and the chairman of the Centrum Rubenianum in Antwerp. He is also working on the Corpus Rubenianum, the definitive catalogue raisonné for the Flemish Baroque painter.
Büttner is in favor of the use of A.I. for authentication, so much so that last year he co-authored a paper with Popovici about its application to verify a painting attributed to Anthony van Dyck. In a phone interview, he claimed that the Van Dyck research project was “proof that [A.I.] works.”
Despite this, he is not convinced by A.I.’s judgement on the The Bath of Diana. He provided a 2016 condition report for the same work by the British art historian Gregory Martin, reviewed by Artnet News, that details the case for ruling out an attribution to Rubens. These findings were confirmed by his own firsthand inspection of the work in 2023 with a U.V. light. Among the many reasons for excluding the work from Rubens’s oeuvre are the use of a reddish canvas primer and an underdrawing, both inconsistent with Rubens’s work from the time, as well as the inferior quality of the painting.
As a possible explanation for why A.I. faltered this time, Büttner suggested that the dataset on which it had been trained may have been insufficient. Popovici confirmed that the analysis was made in 2023, before a collaboration with Büttner greatly improved the scope and quality of Art Recognition’s dataset for Rubens.
In 2021, Art Recognition also made the highly controversial assertion that Rubens’s Samson and Delilah, owned by London’s National Gallery, was a fake. Photo: Johnny Green, PA Images via Getty Images.
Büttner suggests that incorporating another known version of The Bath of Diana in the dataset—specifically a fragment housed at the Museum Boijmans Van Beuningen in Rotterdam, which is the only known version that can be attributed to Rubens’s studio—could have enhanced the A.I.’s accuracy. Though this fragment is not part of the lost original Rubens painting, it is the best version that experts know of and is attributed to Rubens’s studio. It therefore provides a useful counterexample to other versions of the work that are not believed by experts to have been produced by either Rubens or his studio, such as the French The Bath of Diana.
Carina said that the Rotterdam fragment was initially included in the training dataset as an example of an authentic work but was later removed due to Gregory Martin’s 2016 assessment that, while it is from Rubens’s studio, it does not bear any evidence of the hand of Rubens. She claimed, however, that an analysis of the French painting was done using both an A.I. trained on the Rotterdam fragment and one that hadn’t been, and that the results were the same in each case.
“A single image among hundreds in an A.I. model is usually too little to impact the outcome,” she explained, “even if it has a high similarity to the prediction image.” It is not clear whether the dataset included any of the 30 known copies of The Bath of Diana that have been identified by the Rubenianum as examples of fakes. According to Popovici, the owner of the painting was provided with the details of the A.I. training and assessment but did not wish to publish them publicly.
The conversation sparked by Art Recognition’s analysis of the French The Bath of Diana shows that, while support for the use of A.I. for authentication is growing among some scholars and even at the periphery of the art market, the process comes with inevitable complications. As Popovici herself advocates, these A.I. tools must be developed in close collaboration with art specialists who have the knowledge and access necessary to develop clean and comprehensive datasets and are also able to identify key use cases for the trained models.
“I am deeply convinced that Art Recognition can do very good things with the right dataset,” said Büttner, who insisted that the case of the French The Bath of Diana is a bad example of A.I.’s capabilities.
As the field of A.I. authentication is a rapidly developing but still nascent one, its findings may need to be called into question or require an update. For example, it may be the case that an A.I. trained on Rubens works in 2023 may be trumped in accuracy by an A.I. retrained in 2025 on a Rubens dataset developed by a leading authority like Büttner.
In 2021, Art Recognition made the highly controversial assertion that Rubens’s Samson and Delilah, owned by London’s National Gallery, was a fake. The judgement was disputed by leading scholars, including Büttner. This case is covered in greater detail in my book A.I. and the Art Market, published by Lund Humphries in the U.S. last month
Source: https://news.artnet.com/art-world/rubens-attribution-ai-authentication-2622500
