Beyond the Basics: AI and Art Authentication

In her essay, “Unmasking Art Forgery: Scientific Approaches” (in The Palgrave Handbook on Art Crime, 2019), Robyn Sloggett outlines the historical use of scientific and technological aids for analysis. In the 1920s, institutions such as the British Museums Research Laboratories and the Harvard Fogg Museum’s Department of Technical Studies supported chemical analysis of pigments. By the late 1940s, forensic analysis was seen as critical to authentication. Specialists such as Giovanni Morelli and Edmond Locard argued that traditional modes of connoisseurship should expand to include rigorous scientific verification, especially in cases where provenance was less stable. In recent decades, this has included technologies such as scanning electron microscopy, x-ray fluorescence, Raman spectroscopy, and radiocarbon dating. In all such applications, scientific data never serves as proof alone but is instead recognized as one of many tools in the authenticator’s arsenal.

AI authentication must be viewed in much the same way—as a tool and not a solution. Companies such as Art Recognition, Hephaestus Analytical, and Vasarik offer AI models that utilize convolutional neural networks (CNNs), an AI programming structure that supports image recognition. This same technology has widespread application from social media’s ability to suggest identity tags to autonomous vehicles that respond to road signs, pedestrians, and potential obstacles. In art authentication, CNNs perform “computer vision,” which goes beyond simple recognition and allows the system to interpret meaningful information, patterns, and systemic differences in a provided data set.

Many AI models for authentication utilize vision transformers (ViTs), which supplement CNN-derived findings. By splitting inputted images into gridded patches, ViTs allow for even closer analysis. Many purveyors of AI authentication produce “heat maps” identifying areas within a single work that match with or defer from patterns in the artist’s oeuvre. To showcase the strengths of their services, Art Recognition has produced studies of known forgeries by the likes of Tom Keating and Wolfgang Beltracchi. Resulting heat maps highlight areas of concern that differ from the known characteristics of an artist’s practice. 

Like generative AI—the broad term for applications such as Chat GPT— ViTs and CNNs are deep learning mod-els. They mimic the human brain’s ability to interpret and make decisions, unlike simpler machine-learning AI which performs a single task based on the input algorithm. As with all AI, performance ability and reliability are only as strong as the data from which the program “learns.”

Datasets allow the machine to recognize the work of an artist under examination, but the number of images necessary can vary widely. Hephaestus claims their model can “learn” from as few as 30 images, while Art Recognition requires several hundred. In most cases, the AI requires both positive identifications derived from an existing catalogue raisonné and a control set of fake works, comprised of AI-generated works “in the style of” the artist and work by close stylistic contemporaries. Many artists present unique problems in creating datasets. Most successful uses of AI for authentication have examined artists for whom there is a single, well-established catalogue raisonné. In the case of someone like Jean-Michel Basquiat, for whom there is no catalogue raisonné and for whom traditional authentication has been complicated by legal challenges and the shuttering of authentication services by the artist’s foundation, creating a reliable dataset is almost impossible.

Even when an artist is well-suited to potential analysis, competing AI models have produced conflicting results. In early 2023, researchers at the University of Nottingham and the University of Bradford in the U.K. claimed that the long-controversial de Brécy Tondo painting, sometimes attributed to Raphael, was in fact an original work by the artist, reporting 95 percent resemblance to key stylistic attributes. But later that year, Art Recognition’s model suggested an 85 percent probability that the work was not by Raphael.

There are various other challenges facing the new technologies. Machine learning-based analysis tends to have difficulty when an artist’s oeuvre demonstrates stylistic shifts or inconsistencies. Furthermore, little work has been done to develop AI authentication tactics for work in media other than painting. Abstract artists present their own challenges, especially when their work does not exhibit gestural, signatory mark-making. Even so, several recent studies have reported success in utilizing AI to determine the authenticity of work by Jackson Pollock.

Based on current technologies, AI can be best applied in three areas: In performing stylistic analysis, AI can successfully recognize patterns and nuance in an artist’s methods, including brush strokes, color palette, and repeated compositional tactics. In one recent study, AI studied portrait poses in of hundreds of artworks, suggesting significant differences in head position (pitch, yaw, and roll) among the sample set and was able to accurately determine whether works were from Japanese ukiyo-e or modern neo-primitivist artists.

AI can also be well implemented in material analysis, comparing features such as pigment types, binders, and canvas texture to ensure that a work fits within both the time period and the specific characteristics of an artist’s practice. Finally, experts agree that AI can provide much-needed support for provenance tracking. Because of its capacity to quickly and efficiently search and consider large data sets, AI may help save time in the review of historical records, auction catalogues, and other archival records. 

In the arena of provenance and ownership, Non-Fungible Tokens, or NFTs, are touted as a potential deterrent to forgers and provide records of creation and ownership. Linked to cryptocurrency, NFTs can protect against unsanctioned copies and problems of ownership that have plagued digital media. When an NFT is sold, that transaction is documented on the blockchain, a purportedly inalterable archive of crypto transactions. Collectors of contemporary art may be most familiar with the rise of NFTs around 2021. In March of that year, digital artist Beeple’s Everydays: The First 5000 Days sold for over $69 million at Christie’s.

Several recent cases show the ways in which records of ownership on the blockchain can both disrupt traditional systems of ownership and provide a valuable resource for provenance research and disputes over authenticity. In 2021, two companies—Sygnum, a digital asset bank, and Artemundi, a non-traditional investment firm—joined forces to create Art Security Tokens (ASTs) for Picasso’s oil on canvas Fillette au béret (1964). Tokenization allowed multiple patrons to co-own the work. By the close of the subscription period, 50 investors had bought into the painting, creating a first instance of digitally authenticated co-ownership of a major work of art.

That same year, Damien Hirst offered 10,000 works for sale known collectively as “The Currency.” Purchasers had one year to determine whether they wanted to own either the NFT or a related unique work on paper; unclaimed works on paper were allegedly burned. According to the artist, the project challenges systems of value within the market and by extension poses questions about what an authentic work of art is or can be.

Digital and AI-driven authentication practices are here to stay. Major institutions such as the Van Gogh Museum and the Louvre have made recent investments in AI to confirm historically difficult works in their collections. Private entities and public scholars alike continue to document the technology’s capacities. Nonetheless, as Hephaestus CEO Denis Moiseev has argued, AI is “not a silver bullet.” As the digital humanities scholars Peter Bell and Fabian Offert argue, sole reliance on AI authentication will precipitate a new kind of formalism of the type that past art historians, such as the 19th-century Viennese expert Alois Riegl, rejected. Best practices in the field require a balance between formal analysis—a skill set at which AI excels—and contextual analysis—a learned and practiced skill of which only human specialists are (as of now) capable.