AI meets Old Masters in the fight to authenticate paintings

Artificial intelligence is making its mark on the art world, encroaching even on fustier areas such as the Old Masters trade. AI will, for instance, be a talking point during the Tefaf art and antiques fair in Maastricht next week: Carina Popovici, chief executive of the Swiss-based AI company Art Recognition, will reveal at the accompanying Art Business Conference how it recently attributed a painting to a Renaissance German artist.

Art Recognition, which was founded five years ago, has an AI system which, it says, “offers a precise and objective authenticity evaluation of an artwork”. On its website, the company says it has completed more than 500 authenticity evaluations, verifying contested works such as an 1889 self-portrait by Vincent van Gogh at the National Museum in Oslo.

Attributions matter in the art world: confirming the authorship of a work can increase the price if the artist is a star name, and can also boost scholarship in the field. “The Adoration of the Kings”, offered at auction in 2021 with an estimate of €10,500-€16,000 as Circle of Rembrandt, was later attributed to the Dutch master himself and sold for £10.9mn with fees at Sotheby’s in December. And it is not only collectors and dealers who want queries settled: Popovici says Art Recognition is used by wealth management services and legal professionals as well. Christie’s says it is watching “developments in the area of AI with interest”

AI is excellent at so-called “pattern recognition”, says Jo Lawson-Tancred, author of the forthcoming publication AI and the Art Market, so it will have an easier time than humans learning distinguishing features if shown enough examples by a particular artist. It can usually flag any paintings that do not fit an artist’s pattern — but AI does not excel at grasping context, so human reasoning is still integral, she adds.

“I think that a lot depends on the data that is fed into the AI system,” says Carlo Milano of Callisto Fine Arts, London. “For example, if a questionable catalogue raisonné is used to input data about an artist, then the conclusions can be questionable.” The work of an art dealer involves extensive psychology, he explains, outlining that AI will provide more information and reduce the margin of error, but will never completely replace hands-on experience.

Conservators are concerned whether AI can take into account factors such as a filthy layer of varnish, wear or damage. Art professionals are indeed largely sceptical about whether AI will ever supplement or replace the human eye in judging a work of art.

Art Recognition was caught up in a row last year over a painting known as the de Brécy Tondo, believed to be by the Renaissance master Raphael. In January 2023, an analysis by two UK universities (Bradford and Nottingham), using AI-assisted facial-recognition software, concluded that the faces in the work were identical to those in another Raphael painting, the Sistine Madonna (c1513), thus claiming that the de Brécy Tondo was by the master. However, Art Recognition also analysed the piece, by contrast stating that de Brécy Tondo is not by Raphael, with an 85 per cent probability rating.

The Raphael dispute has since broadened, highlighting the strengths and weaknesses of AI authentication — different programs can produce different results. In December 2023, a team led by scientists from the University of Bradford presented further findings about the art of Raphael in a peer-reviewed paper published in the Heritage Science journal. Their program compared the details of authentic Raphael paintings from the database with the test image, examining in depth the colour palette, tonal values, hues and brushstroke patterns.

They concluded that the face of Joseph in the artist’s work “Madonna della rosa”, housed at the Prado in Madrid, may not be by the Renaissance artist. The rest of the work is by his hand, say the university specialists, who include Prof Hassan Ugail, director of Bradford’s Centre for Visual Computing. Ugail says his most recent algorithm recognises authentic works by Raphael with 98 per cent accuracy.

But Popovici challenges Ugail’s findings after investigating the training data made publicly available by the university research group. There are no “negative examples” of Madonna paintings — works not by Raphael but resembling his style — used by the group in its recent Raphael data set, she tells the Financial Times. The validity and scope of the data sets used in such AI programs — the material the software is working from — are crucial.

Like an expert learning from examples, says Popovici, an AI’s ability to recognise patterns and make assessments depends heavily on how representative its training data is, she says. “Without exposure to both authentic and imitative examples of a theme, an AI is inclined to classify as authentic those works that resemble the images in its positive training set,” she adds.

Ugail says that there are other ways to train an AI program than the one Popovici suggests, but strikes a conciliatory note, addressing wider concerns about the impact of AI. This is not a case of AI taking people’s jobs, he says, adding that the process of authenticating a work of art involves looking at many aspects, from its provenance to the pigments used. “Just like spectroscopy and dating techniques, AI can be one important tool in the main toolbox,” he says.

The art historian Bendor Grosvenor says that AI can be useful for connoisseurs. “But the main drawback at the moment is the quality of the inputs given to the AI attribution programs currently being used. It is simply not possible to determine whether a painting is by Rubens by relying only on poor-quality images of not much more than half his oeuvre . . . No human connoisseur would be trusted to do so — neither can a computer.”

His conclusion? “‘Must do better’ is the report card on AI in this field so far.”