Earlier this month, Art Recognition, a Switzerland-based artificial intelligence (AI) company, claimed its software was identified with 96 percent probability that a previously unknown painting was by Italian artist Raphael, an attribution that would raise the value of works to hundreds of millions of dollars. In November, the company’s technology also identified a Titian work held in Switzerland’s Kunsthaus Zurich museum as a likely fake.
Despite the well-publicized revelations, experts in the art world are approaching these claims with skepticism.
Art attributions often shift and move in the museum world, said art advisor Todd Levin, director of the Levin Art Group. It isn’t uncommon to see attributions change as factual knowledge, documentation, scientific testing or new expertise comes to light over the years, he said.
A number of different tools are used for authentication, said Levin. In addition to the judgment of art historians, scientific services such as X-rays and radiation studies can help isolate authenticating factors within artwork. “AI is just another tool, that’s all it is,” said Levin. But it’s a tool that should be used alongside others, he said, instead of a tool that claims to give a final judgment on authentication in return for payment. AI companies are “looking to provide a financialized service that they can make money on, they’re not doing this out of altruism.”
Art Recognition, which was first founded in 2019, claims customers can upload an iPhone quality image of their artwork and receive results on its attribution in as little as seven days. “By telling you what you need to know, we remove any uncertainty on whether an art piece is authentic,” reads its website. The company’s AI technology examines brushstrokes, edges, shapes, color variations and composition elements from its dataset of images to identify the distinctive features of the analyzed artist, said Carina Popovici, Art Recognition’s CEO, in an emailed statement.
Customers include private collectors, auction houses, galleries, dealers and wealth managers, said Popovici. The company charges between 1,000 to 6,000 Swiss francs ($1,060 to $6,370) for its services, which includes either a complete AI report with a detailed explanation of the result or a certificate of the painting’s attribution as authentic or inauthentic. “Artists like Modigliani, with several known catalog raisonnes and little consensus on which artworks are ‘real’ and should be included in the training, are particularly difficult to tackle and therefore more expensive,” she said.
How reliable is the technology itself?
An AI’s analysis will only be as good as the data with which it is programmed, said Karen Thomas, owner of Thomas Art Conservation, a New York-based art restoration company. This becomes more concerning with older paintings which are more likely to have varying conditions of damage. When looking at the bodies of work from a long-dead artist, it’s likely that different paintings will be restored to varying degrees. “I don’t see how you can make a pure comparison between those paintings,” she said.
Despite the new nature of this technology, AI authentication services still provide benefits when it comes to attributions, according to Ahmed Elgammel, a computer science professor at Rutgers University who also founded Artrendex, a company which builds AI services for the art market.
While other authentication services such as canvas or pigment analysis are helpful when it comes to examining the physical properties of artwork, AI technology is focused on stylistic examinations, said Elgammel. Artrendex, for example, examines a work’s brush strokes in order to decide whether they were deliberately chosen by a forger. Authentication at a stylistic level is typically the work of an art history expert, he said, adding that AI performs the same task but has the ability to amass far more relevant data.
However, he still doesn’t believe AI technology should be the sole purveyor of art authentication. While Artrendex undertakes authentication examinations, the company won’t accept works which it believes require other forms of testing. And Artrendex’s final result isn’t marketed as a final conclusion, but a tool which can help aid further human analysis.
“We believe the technology is very new and has big implications, but it can also be misused,” said Elgammel. “There’s no tool that would be valid by itself only.”