Skip to main contentSkip to navigationSkip to navigation
Samson and Delilah, 1609-10, said to be by Peter Paul Rubens in London’s National Gallery
Samson and Delilah, 1609-10, said to be by Peter Paul Rubens in London’s National Gallery. Photograph: agefotostock/Alamy
Samson and Delilah, 1609-10, said to be by Peter Paul Rubens in London’s National Gallery. Photograph: agefotostock/Alamy

Was famed Samson and Delilah really painted by Rubens? No, says AI

This article is more than 2 years old

Long-held doubts about the authenticity of the National Gallery’s masterpiece, bought for £2.5m in 1980, are backed by pioneering technology

The National Gallery has always given pride of place to Peter Paul Rubens’s Samson and Delilah, listing it among the “highlights” of its collection, since it purchased the picture at Christie’s in 1980 for a then record price.

It depicts the Old Testament hero in the lap of the lover who betrayed him, having beguiled him into revealing that his God-given strength lay in his uncut hair. As Samson sleeps, Delilah’s accomplice cuts his locks, rendering him powerless, with soldiers ready at the door to capture him.

Critics have long suggested that the painting is not really by Rubens. And now a series of scientific tests employing groundbreaking AI technology have concluded that the 17th-century Flemish master could never have painted it.

“The results are quite astonishing,” Dr Carina Popovici, the scientist who carried out the study, told the Observer. “The algorithm has returned a 91% probability for the artwork not being authentic.”

In comparing Samson and Delilah with 148 uncontested Rubens paintings, the analysis produced one of the scientists’ highest-ever probabilities of the hundreds of pictures they have tested so far.

“I was so shocked,” said Popovici. “We repeated the experiments to be really sure that we were not making a mistake and the result was always the same. Every patch, every single square, came out as fake, with more than 90% probability.”

The AI analysis was carried out by Art Recognition, a Swiss company based near Zurich of which Popovici is co-founder. It has analysed 400 works with this technology, and has an ongoing collaboration with Tilburg University in Holland.

Its report concludes: “The AI System evaluates Samson and Delilah not to be an original artwork by Rubens with a probability of 91.78%.” In contrast, the scientists’ analysis of another National Gallery Rubens – A View of Het Steen in the Early Morning – came out with a probability of 98.76% in favour of the artist.

The National Gallery bought the picture for £2.5m, equivalent to £6.6m today, and some have suggested that it would now fetch far more as The Massacre of the Innocents sold for a record £49.5m in 2002.

Michael Daley, director of ArtWatch UK, who has researched the painting extensively, described the AI report as “exceedingly damning”. Daley is among those whose serious doubts about the painting’s attribution to Rubens have always been dismissed by the gallery and Rubens scholars.

Critics have long argued that it is only a copy of a Rubens original that is known to have been painted between 1608 and 1609 for his Antwerp patron Nicolaas Rockox which then disappeared after his death in 1640.

They argue that the National Gallery picture is a different painting, one that only surfaced in 1929, declared a Rubens by Ludwig Burchard, an expert who, after his death in 1960, was found to have misattributed paintings by giving out certificates of authenticity for commercial gain.

The picture’s critics dismiss its colours as uncharacteristic of Rubens’s palette and its composition as awkward. They question why, for example, it differs from two contemporary copies made from Rubens’s original. The toes of Samson’s outstretched right foot, for example, are cropped in the National Gallery version, while they are shown in an engraving by Jacob Matham and a painting that depicts the Samson and Delilah hanging in Rockox’s home by Frans Francken the Younger.

“For three decades, the National Gallery dismissed artistically, technically and historically informed challenges to the attribution. When a picture is wrong, everything is wrong,” Daley said.

The National Gallery has faced similar criticism in the past. It included the Salvator Mundi in its 2011 Leonardo exhibition, although it was an unknown work with doubts about its attribution, restoration and ownership. Its display would have boosted its market value and the painting sold at Christie’s, New York, in 2017 for $450 million. “Coming so soon after its ill-advised espousal of the now-rejected $450m Salvator Mundi, these Rubens findings are a calamity for the National Gallery,” Daley added.

Dr Katarzyna Krzyżagórska-Pisarek, an art historian, has long dismissed the Samson and Delilah as “highly problematic”, having identified more than 60 of Burchard’s Rubens attributions that have subsequently been demoted.

“The significance of this new AI method of authentication is potentially groundbreaking. Devoid of human subjectivity, emotion and commercial interests, the software is coldly objective and scientifically accurate. Many questionable works were attributed to Rubens at the beginning of the 20th century… There is today a distinct need for more reliable methods of connoisseurship,” she said.

The AI technology – “convolutional neural network” –analyses different features of an artist’s work, including brushstroke patterns. Popovici said that “everything that’s not by the hand of the assumed artist will come out with a negative probability” and that the technology can produce accurate results even for artists who significantly changed their style across their careers as every artist has a unique brushstroke.

She added: “Sometimes patches don’t come out with the same probability, if it’s not clear but, in this case, every single patch came out with more than 90% probability as a fake. This is very certain.”

A National Gallery spokesman said: “The Gallery always takes note of new research. We await its publication in full so that any evidence can be properly assessed. Until such time, it will not be possible to comment further.”

This article was amended on 26 September 2021. An earlier version referred to a “convulsional” rather than a “convolutional neural network”.

Most viewed

Most viewed