> TELEPHONE US 212.355.5710

The Achilles Heel Is Samson’s Foot: Artificial Intelligence In Art Authentication

October 8, 2021
figures in an interior

Samson and Delilah

When it comes to the future of many jobs, one of the most common fears is the advent of automation. The image that often comes to mind is that of robotic arms on an assembly line, repeating the same action continuously without any sign of fatigue. But the manufacturing sector is not the only place where automation may threaten peoples’ job security. Of all the organizations interviewed by McKinsey, 31% are worried about labor displacement because of automation. Because many see automation as an inevitability, there have been efforts to introduce artificial intelligence as a replacement for a great number of jobs, some of which actually require humans. The most recent example I saw is several different companies coming out with their own countertop bartending machines, dispensing Manhattans and Negronis with nothing more than a disposable pod and the push of a button. It seems like the same is being done with art authentication.

Art authentication software got a big win not long ago, when the Guardian reported on September 26th that one of the most prized paintings in London’s National Gallery has been declared a forgery by an art authentication AI. Samson & Delilah, attributed to the Flemish master Peter Paul Rubens, has been kept at the National Gallery since 1980, when the museum bought it at a Christie’s auction for £2.5 million (nearly £11 million today). Using 148 verified Rubens paintings as a control group, AI software from the Swiss-based tech company Art Recognition reported that there is a 91% probability that Samson & Delilah is a fake.

The AI used by Art Recognition is what is known as a convolutional neural network, which divides a painting into a grid, then analyzes each cell to detect color and brushstrokes patterns. Despite some sources hailing this new technology as a sort of miracle machine ready to replace the actual human experts, the people behind the software have emphasized that this is not the case. Dr. Carina Popovici, one of the co-founders of Art Recognition, has made it clear that the software, while robust enough to detect fakes and forgeries, is best used in tandem with “[c]hemical analysis of pigments or carbon detection” to better aid experts in authentications.

But when one looks at the history of the National Gallery Rubens, it seems like the AI did nothing but tell everyone what most experts already knew. According to records, there was indeed a Samson and Delilah painting created by Rubens around 1610 for the mayor of Antwerp, Nicolaas Rockox. That work was sold to an unknown buyer in 1640 and disappeared from the record not long after. The work in the National Gallery first appeared when the Rubens expert Ludwig Burchard authenticated it in 1929. After Burchard’s death, however, it was revealed that he had knowingly issued certificates of authenticity for fake Rubens paintings in exchange for cash. This revelation threw many of his authentications into doubt, including the one for Samson & Delilah. Furthermore, copies and engravings of the original painting show several major differences between the National Gallery Rubens and what the original should look like. One difference may seem insignificant, but it is one of the most frequently-cited pieces of evidence used by skeptics: a foot.

In the painting, Samson lays asleep in the lap of his lover Delilah as she cuts his hair. Meanwhile, his leg extends to the right-hand side of the canvas until his toes disappear off the edge, out of view. Some Rubens experts have drawn attention to this as evidence that the National Gallery painting is a copy. Human feet are notoriously difficult to draw and paint. Case in point, even the Italian Renaissance master Sandro Botticelli was known for never being able to master the art of painting feet, and consequently left them rather awkward and misshapen in his finished works. Copies and engravings of the original Samson & Delilah show the full foot included in the painting, for good reason. Because feet were so difficult to depict, Rubens made a point to feature them in their full form in many of his works, including Descent from the Cross, The Fall of Man, David Slaying Goliath, and the Three Graces. Despite the evidence accumulated even before the AI analysis, the National Gallery and some Rubens experts continued to assert the work’s authenticity. Details like Samson’s foot may be something that AI might skip over while looking at brushstroke patterns. There are also other factors that make the AI report less significant. The Art Recognition software was programmed only to recognize the characteristics of a Rubens work based on shared traits from a cache of other Rubens works. But the AI was not designed to recognize the characteristics of his studio apprentices that aided in the completion of his works. In the past, some works attributed to Rubens were later identified as created by Jan van den Hoecke, Rubens’s main assistant in the 1630s.

The only response from the National Gallery was a brief recognition of the report: “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.” While AI software should definitely be considered a tool that experts can use in the authentication of artworks, they shouldn’t be thought of as a total replacement. There are some jobs that machines can do very efficiently, but then there are some that require an actual human touch. Which brings me back to those countertop bartender machines. But does that little machine really replace an actual bartender? Can it make recommendations based on your palate, or make witty jokes as it strains your cocktail into a coupe glass, or come up with new drinks based on their own experimentation? While the data collected by AI can be useful, identifying a painting as a forgery involves a lot more than that. Knowing the provenance, getting a look at the back of the canvas, running chemical tests on the pigments, and being intimately familiar with an artist’s entire oeuvre is required to make that kind of judgment. There’s even a healthy bit of instinct involved, or realizing that there’s something that’s not quite right without being able to specify the issue. Until someone can develop artificial intelligence that can do all that, all the experts can rest easy for now.