Masterpiece or cheap copy? Art historians and AI may not agree

By Oscar Holland, CNN

To the untrained eye, there is very little difference between the three known versions of “The Lute Player.” Almost identical in composition, the paintings all depict a young, doe-eyed subject in white robes, instrument in hand and turned slightly away from the viewer. Each appears to carry Italian painter Caravaggio’s signature mastery of light and shadow.

To art historians, however, there has long been broad agreement: The versions held by Russia’s Hermitage Museum and France’s Wildenstein Collection were created by the Baroque artist, while the one at Britain’s Badminton House is merely a copy.

Artificial intelligence begged to differ. In September, Swiss AI firm Art Recognition claimed there is an almost 86% chance that Badminton House’s version is, in fact, authentic. The company’s model, which was trained to recognize markers of Caravaggio’s style, including shapes, color palettes and compositional structures, also declared (albeit with less statistical certainty) that Wildenstein’s version is likely a copy. Its analysis found a “significant divergence” between the latter painting’s “visual characteristics” and those of Caravaggio’s other works.

This is one of several bold claims made by Art Recognition since it launched seven years ago. In 2021, the company calculated a 91% chance that a painting at London’s National Gallery attributed to Peter Paul Rubens, “Samson and Delilah,” was not produced by the Baroque painter. A long-disputed painting of Vincent van Gogh at the The National Museum in Oslo, meanwhile, had a 97% chance of being genuine. The firm’s other analyses have presented more complex results: Rembrandt’s “The Polish Rider,” for instance, was partly produced by someone else, though some sections carry evidence of the Dutch painter’s hand, ranging in certainty from 69% to 83%, according to the AI model.

Art Recognition’s declarations have not always contradicted the established scholarship. The Van Gogh attribution, for instance, was subsequently matched by more conventional research, including technical analyses and studies of the artist’s letters (museum experts concluded that the portrait’s unusually dampened colors simply reflected Van Gogh’s troubled mental state at the time). Yet, many art experts remain highly skeptical about AI’s ability to supersede, or even complement, the tools traditionally used to authenticate works of art.

“I think it’s quite problematic,” said Angelamaria Aceto, a senior researcher at the University of Oxford’s Ashmolean Museum of Art & Archeology. “I’m very open to new technologies; I use technologies all the time that can help you to see what the naked eye can’t — to go beneath the surface. And I’m sure AI is fantastic at analyzing data and providing data, but connoisseurship is about contextualizing things. It’s about thinking critically.

“I may go to a conservation scientist and ask them to analyze a pigment; I may ask a photographer for an infrared image,” she added. “But thinking AI can substitute the educated, critical eye? That’s a no-no for me.”

Seeing what humans can’t

Combining machine learning, deep neural networks and computer vision algorithms, Art Recognition’s approach can, in theory, be adapted to any painter with a big enough back catalog. To date, the company has produced models for more than 200 artists.

In each, the AI is trained on two photographic datasets: A “positive” one, containing images of undisputed (or widely accepted) paintings by the artist in question, and a “negative” one comprising similar, but inauthentic, works. The latter group may include known forgeries, copies by students or admirers — like Caravaggio’s 17th-century stylistic followers, known as the “Caravaggisti” — and even AI-generated images, created in the artist’s style.

Having a “high degree of similarity” between the two datasets is crucial, said Art Recognition’s co-founder and CEO, Carina Popovici, on a call from Switzerland. “We really want the AI to learn the difference between Caravaggio and an imitator of Caravaggio — the difference between a Rubens and an almost-identical painting created in his workshop by an apprentice.”

To prevent bias, both training datasets feature a comparable balance of subjects and genres, such as portraits, landscapes, still lifes or religious scenes. Developers include augmented versions of each image, mimicking different lighting conditions and flipping or rotating high-resolution photos of the paintings, to expose the models to different spatial configurations. Images are also divided into small squares, or “patches,” that force AI to consider the artwork’s characteristics in new ways. Looking at a smaller, quieter part of a painting in isolation may help it learn fine brushstrokes, for instance, while a zoomed-out view could teach it about composition or color.

But Popovici admits it is not always clear how the models reach their conclusions. “We can speculate, but we don’t really, really know for sure,” she said, adding there will be “some types of patterns that AI can see better than humans.”

It takes up to a week to train the AI on the chosen artist, then another day or two to analyze an individual artwork. However, Popovici said the most time-consuming part is researching and building the datasets. Like any AI model, the output is only good as the data it is trained on (or “garbage in, garbage out,” as the saying goes).

For this reason, Popovici argues that Art Recognition does not sit apart from art history — it relies on it. The firm’s datasets are, she said, “the product of scholarly expertise,” built by in-house historians who have studied artists’ biographies, catalogs and academic literature. She hopes AI can be a “tool in the toolbox” for experts, not a replacement for them.

“I think it’s very counterproductive to be in a perpetual argument about who’s right, the expert or the AI?” she said. “We don’t want to be their enemy.”

More than surface deep

The experts CNN spoke to were, however, less conciliatory about a technology that expresses conclusions without explaining how it reached them.

“Authentication is rarely just about surface style,” said art historian and curator Sharon Hecker. “It involves a lot of art-historical context. You have to know about the workshop, practices, materials, the condition of the work, the restoration history and how a particular artist’s work evolved over time.

“If you think about patterns of brushstrokes, who’s to say an artist didn’t change styles one day — they woke up and worked in a different style?” she added. “Artists are just not that predictable, and that unpredictability is part of the beauty of art. So, AI being trained to recognize a consistent style may miss a lot of these nuances.”

Among Hecker’s primary concerns about operations like Art Recognition (whose customers frequently include owners looking to sell under-researched paintings inherited from deceased relatives) is their lack of transparency. Without full access to the companies’ commercially sensitive models and datasets, researchers are unable to check AI’s workings, she said. “Any science has to be independently replicable,” Hecker added. “I would have to be able to replicate the study, using the exact same dataset, and come back with the same result.”

AI studies may not always even agree with one another. In 2023, a disputed Raphael painting, the “de Brécy Tondo Madonna,” was subjected to two analyses that came to very different conclusions. One, by researchers at the UK’s University of Bradford, used AI-assisted facial recognition to compare the painting to Raphael’s “Sistine Madonna,” finding both were “undoubtedly” by the same artist. Art Recognition’s model meanwhile found an 85% chance that the painting was not by the Renaissance artist.

In response to the resulting debate, then dubbed “the battle of the AIs,” Art Recognition argued that finding facial similarity between two Renaissance paintings was “hardly surprising” and should not constitute an attribution claim. The company then made its Raphael dataset public, saying transparency was “not a public relations gesture” but a “prerequisite for credibility.”

A matter of trust

In the art market, credibility is everything. As such, the question facing AI researchers is not just whether their methodologies are sound — it’s whether anyone is willing to believe, or even listen, to them.

Determining if a painting is “authentic” is often a matter of broad, rather than absolute, consensus. Historical paintings don’t come with certificates; experts are increasingly hesitant to express complete certainty, for fear of litigation or reputational damage should they be proven wrong. For similar reasons, many of the foundations and artists’ estates once considered the ultimate authorities (like the Keith Haring Foundation and the Andy Warhol Foundation for the Visual Arts) have ceased offering authentication services.

“I think there’s a problem in the word ‘authentication,’” said Hecker. “People really want a level of certainty that probably doesn’t exist, whether it’s a machine or the human eye.”

Instead, buyers must rely on trust — in museums, auction houses and the market itself. After all, someone’s willingness to spend millions of dollars on a painting can speak volumes, given that a downgraded attribution can knock numerous zeroes off an artwork’s value.

For Art Recognition, however, a landmark moment arrived in late 2024 when a Swiss auction house used the company’s research to back the sale of three artworks. Among them was a watercolor, ostensibly by Russian artist Marianne von Werefkin, that was otherwise short of evidence supporting the attribution, according to Popovici.

“It was a big moment, because it showed that this is not just an academic tool, but the kind of product that can really make an impact on the market,” she said, calling the auction a “turning point.”

Whether AI research will be embraced by bigger auction houses, or used to back more valuable sales (the Von Werefkin watercolor fetched a modest 15,000 Swiss francs, or $19,600), is another matter altogether. So, too is the question of whether museums will ever take the findings seriously. Although Art Recognition has collaborated with institutions like the Kunsthaus Zürich, Popovici concedes that galleries have little incentive to embrace technology that might cast doubt over their collections.

She said London’s National Gallery “didn’t want to talk” to her about the firm’s findings on “Samson and Delilah.” In the years since, further evidence questioning the attribution has emerged, though the museum has consistently defended its longstanding attribution to Rubens. In a statement emailed to CNN, the National Gallery spokesperson said that an “extensive study, conducted by our curatorial and scientific teams using the latest imaging and analytical techniques, provides compelling evidence in support of the painting’s authorship,” adding: “Not one single Rubens specialist has doubted that the picture is by Rubens.”

Yet, despite experts’ distrust, might AI at least help start conversations, even if it’s not their final word? Popovici expressed hope that her company’s research could help “unlock” paintings that are “otherwise just sitting in a basement somewhere.” And art historian Hecker acknowledged that AI “could flag an issue” that is then investigated using conventional research, caveating that she “would be more comfortable using a university-based laboratory that doesn’t have a large commercial interest.”

The technology might also be used to expose criminal activity or help vendors weed out fakes, argued Popovici, who was inspired to start Art Recognition after learning about the case of German master forger Wolfgang Beltracchi, who was imprisoned in 2012 for a multi-million-dollar fraud that fooled buyers, galleries and auctioneers.

“Experts were checking every single one of those paintings, and they all gave the green light,” she said, arguing that the human eye “inherently makes mistakes.” Popovici has since used the company’s models to analyze numerous Beltracchi paintings, which were created in the style of various deceased European artists, to see if they would have successfully exposed his deception.

“Everything came out as fake,” she said. “And with very high probabilities.”

The-CNN-Wire
™ & © 2026 Cable News Network, Inc., a Warner Bros. Discovery Company. All rights reserved.