AI Is Trying to Take Over Art Authentication, but Longtime Experts Are Skeptical

There’s something dystopian about a tech company cofounder standing on stage and criticizing humans for being slow, expensive, and outdated in authenticating art. Why send a painting to some crusty old art expert’s laboratory for subjective analysis when “objective” artificial intelligence can do the job faster and more cheaply using just photos?

That was the question posed by Carina Popovici, CEO and cofounder of Art Recognition, a Swiss firm that uses AI to authenticate art, during a TEDxNuremberg talk in early 2022. The moment recalled the 1987 sci-fi blockbuster Robocop, specifically, the scene where an executive of evil mega-corporation OmniCorp unveils its latest police robot to a wide-eyed boardroom. Triumphantly, he tells the room that they need a cop “who doesn’t eat or sleep.” The robot stomps in before malfunctioning and pumping a suited board member full of hot lead. Art Recognition may be not OmniCorp—and Popovici nothing like her fictional corporate counterpart—but the company and art authentication outfits like it are similarly banking on technology to “clean up” the art market of fakes and forgeries. They’re also planning to do it with unprecedented efficiency and automation.

Related Articles

Artists and Creatives Are Working with AI Companies, but Should They?

More Art School Classes Are Teaching AI This Fall Despite Ethical Concerns and Ongoing Lawsuits

If you used a human “you would have to pack your painting, ship it off to a different country for appraisal … then you would have to wait for some months, or sometimes even years, for an answer,” Popovici said, with apparent disdain, as she live-demonstrated Art Recognition’s tech. “Our program needs about three days to learn the characteristics from around 700 training images, and less than five minutes to calculate the probability of the authenticity of an artwork.”

Art Recognition is far from the only company leveraging AI for art authentication, which has become one of the most popular use cases for the technology in the art world.

Hephaestus Analytical is a London-based tech company that integrates AI analysis and machine learning trained from sampled data sets, alongside scientific tests, provenance research, and “connoisseurly expertise to analyze works. It is focused on arguably the “dirtiest” corner of the market, the Russian avant-garde, which also includes modernism that flourished in other Soviet nations during the late 19th and early 20th centuries. Denis Moiseev, the founder and CEO of Hephaestus, told ARTnews that more than 95 percent of the Russian avant-garde paintings brought to him are fake. (One London-based dealer specializing in Ukrainian modernist artists, James Butterwick, told ARTnews something similar, claiming that as much as 95 percent—“in fact, probably more”—of the paintings offered him are not authentic.) Hephaestus claims its system produces “the most conclusive authenticity results.”

“The market is so saturated with forgeries, however, that it’s impossible not to talk about it,” Moiseev said. “We believe our technology can clean up the market. It is a solvable problem. The issue is that there’s an adversarial component to the Russian avant-garde market—there are so-called experts who are authenticating, or contributing to the authentication—of forgeries. There are people who say things are real and actually, they’re not. This is what makes this market so complicated, but it really shouldn’t be.”

Moiseev said that he understands some people might find Hephaestus’s data-driven approach “cold-hearted” but he is on a mission to “find the unique features, characteristics, and data points that make artists unique.” Moiseev believes the art market has “benefitted from ambiguity” but now needs to “open up” and accept that “scientific analysis and technology have a role to play in authentication.”

“At the moment, that’s not what we’re seeing,” he said. “We’re seeing … huge … reticence from the art market to adopt these technologies.”

He pointed to auction houses closing their scientific research departments or placing them “on demand” as evidence that the art market views science as a method of “last resort.”

“This is a big problem,” Moiseev said.

While Sotheby’s declined to comment on the use of AI in art authentication, it did tell ARTnews that though its scientific research department has had “a period of inactivity,” it remains “operational” and a “valuable resource.” A Christie’s spokesperson told ARTnews that the company is exploring how AI solutions can “enhance our productivity and efficiency.”

“Our business is constantly evolving and embracing new tools for innovation to support our ability to provide the best service to our clients,” the spokesperson said. “AI is no exception, and we see the value it could add. We believe this is about augmenting intelligence as no digital tool will ever replace the passionate expertise or trusted close relationships Christie’s is proud to share with our clients.”

Nicholas Eastaugh, the CEO of Vasarik, another London-based AI art authentication company, is optimistic about the role humans will continue to play in the field.

“This should not be seen as a process of either/or with regards to AI replacing human judgment, but one of AI providing tools that experts can use,” Eastaugh told ARTnews. “Currently, the weakest link I see is in the sets of images used to train AIs. Frequently, these are badly chosen, failing to reflect the kinds of judgments art experts need to make. Art historical knowledge allows us to ask better questions of the AIs and consequently get better results that can be trusted.”

To Eastaugh, because AI only produces a probability of how likely a painting is to have been created by a particular artist, the results are “always in a sense provisional.” Other information needs to be considered to reach the most reliable results, such as chemical composition. This is how Vasarik works, combining AI analysis with art historical knowledge and scholarship.

Hephaestus’s Moiseev also emphasized that AI is a tool, “not a silver bullet,” and should be used alongside human expertise and scientific testing. In fact, the company’s founding mission was to eliminate forgery and misattribution from the art market with a protocol including chemical analysis, provenance research, and connoisseurship.

“Chemical analysis allowed us to date materials, but not evaluate the likelihood that a given artist produced a picture,” Moiseev said. “Machine learning provided a way of scientifically identifying the unique characteristics in an artist’s work.”

Hephaestus’s AI needs only 30 images of an artist’s work to train its AI to authenticate a painting, a number that Moiseev concedes is “incredibly low,” particularly given that Art Recognition’s AI, by comparison, needs several hundred.

“We work by effectively training algorithms on a set of carefully curated images of 100 percent authentic artworks that have never been questioned. The technology extracts unique features related to brushstrokes, including the variation and curvature of the stroke, which are linked to the characteristics and motor skills of each artist,” he explained. “One way to describe it is like this—scientific analysis looks into the painting or through the painting; provenance research looks at the history; connoisseurship looks at the current picture in front of you; while AI looks across the body of work for these granular details, these clustering of brushstrokes to identify whether something is unique.”

While Popovici of Art Recognition said she was reluctant to comment directly on Hephaestus needing only 30 images to train its AI, she said “any AI developer would tell you that no serious AI can be trained on such a low number of images.” She added that her statement in the TEDx talk about Art Recognition needing 700 images for an artist was “oversimplified.” For artists with more complex or varied oeuvres, the AI might need thousands of images.

“The number of images is a poor unit of measure when it comes to determining the effectiveness of AI,” Moiseev said. “Not all images are equal—there are many factors at play, such as resolution or image quality—and not all algorithms are designed the same. Claiming that quantity trumps quality in data is like saying a crowd of amateurs is more effective than a handful of experts.”

Simon Gillespie, who runs an eponymous art authentication and restoration studio in London, told ARTnews that he thinks of himself as a “top-class surgeon” whose “subtlety of touch” will always be required over AI. He said that while it’s inevitable humans will be substituted for technology in some aspects of authentication, he believes that any company using AI to fully attribute a painting “should be treated with [substantial] doubt.”

“So far, I have not seen any of the AI companies give a 100 percent attribution, this would be very arrogant,” Gillespie said. “But AI can be a very useful tool and it will undoubtedly improve the process.”

Art Recognition, Popovici’s firm, uses a “standalone AI,” in which humans select and curate the dataset, but execute “no human judgment in the actual authenticity evaluation.” Popovici did warn that AI should not be relied upon alone, noting that the results are only as good as the dataset provided, and that there are some cases, as with Amedeo Modigliani, where there are multiple catalogues raisonnés and no consensus on which one is correct.

“In such scenarios, while we are fully transparent about the images we use for training, it is crucial to combine our results with those of human experts,” she said. “I strongly believe that the future of art authentication lies in the collaboration between AI and experts.”

Art Recognition, Popovici explained, provides clients with authenticity probabilities. But when that probability is greater than 95 percent, consulting another expert may not be necessary. When below 80 percent, she often recommends material analysis or evaluations from other experts.

While emphasizing the importance of scientific vetting, Popovici said, “it’s true that not every single detail of our models has been submitted to scientific journals [but] the core components – including very detailed descriptions of the architectures we use – are thoroughly described [in two scientific papers].”

Moiseev said Hephaestus’ “underlying technology” has been peer reviewed, but he seemed more skeptical about the process. “Peer review can sometimes be used as a bit of a smokescreen,” he said. “In the field of art authentication, you are either right or wrong, your results are either accepted by the market, or they are not. At Hephaestus, we’ve seen too many paintings pass external AI examination only to fail basic scientific tests. For us, AI is part of a robust multimodal authentication protocol.”

Bendor Grosvenor, a leading British art historian who has discovered several lost old master paintings, told ARTnews that he is wary of AI.

“I suspect AI will play an increasingly important role in helping us to recognize who painted what, and when,” he said. “At the moment, however, the track record of AI attributions is patchy, to say the least. Probably just as important is the fact that the market is some way from accepting what the computer says, and prefers the judgment of academic research, the human eye, and technical analysis.”

Moiseev, for his part, admitted that, though he has faith in AI authentication technologies, the art world is not close to adopting it as a standalone solution. Hephaestus, he said, is focused on using its technology to make art a commodity of irrefutable provenance so its value can be extracted—that is by working with banks and financial institutions to offer loans against the works.

“We are trying to build an incentive structure,” he said. “Once a painting has passed our protocol and has been confirmed as genuine, for example, you’re dealing with a different object. There is serious value in having no fear of an artwork being inauthentic.”