Two artificial intelligence models were used to authenticate Raphael’s painting and came up with different results, casting doubt on the future of the technology

AI-powered facial recognition technology has identified a previously unattributed de Brécy Tondo painting as the work of Raphael. Photo courtesy of the de Brécy Tondo Trust.

Two different aartificial intelligence models trained to find out if the work known as de Brécy Tondo is Raphael’s hand, expressed two different opinions challenging the growth of art authentication technology.

An AI model developed by Hassan Ugail of the University of Bradford “definitely” recently found that the work, known as de Brécy Tondo is the hand of Raphael. This is now seen for the first time Cartwright Hall Art Gallery UK

“By testing Tondo using this new artificial intelligence model, there are stunning results confirming that it is most likely Raphael,” Ugail told the BBC last month. “Combined with my previous work using facial recognition, and combined with previous research by my fellow academics, we concluded, Tondo and the Sistine Madonna are undoubtedly by the same artist.

Then the model that created Art Recognition – A company that offers technology to authenticate works of art has determined with an 85 percent probability that de Brécy Tondo was not painted by Rafael.

The Swiss company has previously used its technology to test Flaget Madonna as an authentic work Raphael and assert with 92 percent certainty that Samson and Delilah (1609–10) can no after all, to be the hand of the Flemish artist Peter Paul Rubens.

Carina Popovici, CEO of Art Recognition, told Artnet News in an email that she was surprised to find the results of her study “clearly contradict” those of Ugali’s team.

“In art recognition, we train our network on images of an artist’s artwork, ​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​ The Bradford and Nottingham researchers use a neural network that learns facial features from a massive data set of millions of faces, according to their publication, Popovici said.

The authors transferred de Brécy’s Madonna and Raphael images Sistine Madonna through their network, resulting in a high similarity score.

“I’m concerned that this situation could undermine the progress we’ve made over the past five years in establishing AI as a key method for art authentication,” Popovici said. “Now more than ever, it is necessary to emphasize the importance of adhering to rigorous scientific standards. Otherwise, the whole field of AI could come under fire and we would all suffer the consequences.

Italian School, Leonardo da Vinci, copy Salvator Mundi (about 1600). Photo courtesy of Christie’s.

Experts told Artnet News they believe artificial intelligence will never completely replace traditional authentication methods, as the technology’s limitations become increasingly apparent when it is applied to works other than Old Masters.

“Ultimately, there will always be room for human judgment,” art historian Martin Kemp, one of Leonardo da Vinci’s leading authorities., said Artnet News.. “Some traditional historians have been reluctant to accept that machines can tell them about flaws. It’s pretty much being shaken out of the system now.

He sees artificial intelligence as “another tool in our armoury”, such as pigment analysis and X-rays to examine the underlying paintings. But It can be difficult for AI to analyze old masterpieces, e.g Salvator Mundi which are “very damaged,” Kemp added, “and the damage is sometimes different.”

He also said AI could fail when trying to authenticate artists such as Titian or Leonardo, whose style varied from a simpler technique earlier in his career to “fine layers of very thin glaze” later. Raphael, on the other hand, was “much more consistent.”

Martin Kemp, Leonardo expert from Milan Missing Leonardo. Photo courtesy of Sony Pictures.

Larry Silver, an art historian at the University of Pennsylvania who worked on the authentication Flaget Madonnasaid he believed Art Recognition’s authentication data set “performed better” than it could perform human authentication.

“It is an early painting by Raphael and on a small scale. But I’m not sure that the AI ​​training model is nuanced enough to correctly identify the work of an artist whose style has changed over the course of his career,” Silver said, putting Raphael in the same category that Kemp did with Leonardo. .

“I think auction houses, museums and galleries can use AI,” Silver said, adding that “the intervention of a human interpreter is always necessary.”

Popovici said art recognition was designed to reduce the clash of human interpretations and egos while bringing transparency to the authentication process.

She said her first model, which she trained on images of famous Wolfgang Beltracchi forgeries found online, was 100 percent successful in identifying forgeries that weren’t used to train the model.

Peter Paul Rubens, Samson and Delilah (about 1609/10). Collection of the National Gallery, London.

Other artists who could challenge AI include Rubens, Silver said, because he worked with collaborators, especially on larger-scale works. Vermeer is another. “[He] was a very exceptional artist, so it’s impossible to teach artificial intelligence his 30 paintings,” Popovici said. “So, of course, it’s very important to have an expert opinion.”

An even bigger new hurdle facing the authentication space is the growing number of AI-generated images. How can AI authentication deal with a fake old master generated by an algorithm?

“Authentication with AI isn’t really authentication, it’s just styling,” said Ahmed Elgammal, director of the Art and AI Lab at Rutgers University and co-founder of Playform AI. artists”.

He added: “You can tell Monet, Picasso and Van Gogh apart – it’s quite easy. But when you get into a forgery, it determines whether the art is in Van Gogh’s style, and it doesn’t confirm how Van Gogh works.

While AI generators are “probably a long way from physically creating something from scratch on canvas,” Kemp said, advances in technology mean that detailed oil paintings can be printed as they were once painted.

According to Arnold Brooks, a professor of graphic arts at Brooklyn College, printers can now print with oil paints and, for example, can use artificial intelligence to print an underlying image with original or historical pigments.

Elgammal added that counterfeiters “have always been the first” to explore new technologies to make better fakes.

Wolfgang Beltracchi with a forged painting allegedly by Max Ernst. Photo by Brill/ullstein bild via Getty Images.

Using AI to authenticate art raises another question: Could AI help police and law enforcement better identify art seized during investigations?

“AI won’t find something locked in an undefined storage block. At least not,” said Samantha Moore, counsel for the Artists’ Rights Society.

However, AI can show the probability of a painting being stolen. Experts say that opportunities to help law enforcement will become more important as the technology grows in popularity. Popovici said the art recognition organization does not yet have law enforcement cooperation, but has consulted with the Zurich Police Department in Switzerland about one case.

Art Recognition began training its models using artificial intelligence-generated forgeries to train a classifier to distinguish between human and synthetic forgeries.

“An AI-generated image can fool an AI censor,” Elgammal added. “The use of AI for identification and verification is promising, but still in its infancy. Some claims about using AI for authentication can be deceptive.

An FBI spokesperson told Artnet News that the bureau has “always tried to stay abreast of advances in technology that can be used to further a crime or threaten national security.”

“These advances continue to this day, especially in the areas of artificial intelligence and cyber security.

Follow up Artnet News on Facebook:

Want to keep up with the art world? Subscribe to our newsletter for the latest news, eye-opening interviews, and hard-hitting critiques that keep the conversation going.

Godfrey Kemp

"Bacon fanatic. Social media enthusiast. Music practitioner. Internet scholar. Incurable travel advocate. Wannabe web junkie. Coffeeaholic. Alcohol fanatic."