In 2025, a team of researchers published a paper describing a new machine learning method for attributing paintings. They called it PATCH: Pairwise Assignment Training for Classifying Heterogeneity. The method trains a neural network to distinguish between patches of paint applied by different artists working on the same canvas. They tested it on two paintings by El Greco in the Hospital Tavera in Toledo. One, "Christ on the Cross," is entirely by El Greco. The other, "The Baptism of Christ," was begun by El Greco but finished by his workshop after his death in 1614. Art historians had spent decades arguing about which sections were by the master and which were by his son Jorge Manuel or other assistants. The PATCH algorithm analyzed patches of the painting and identified regions where the brushwork was consistent with El Greco and regions where it was not, producing a map of authorship across the canvas surface. Read the paper at arXiv.
Attribution is the process of assigning a work of art to a specific artist. It is the foundational act of art history: before you can study a painting, you need to know who made it. Attribution answers the question "who painted this?" and the answer determines the work's historical significance, its monetary value, and its place in the narrative of art. Attribution is not a one-time event. It is an ongoing process of revision, debate, and reassessment that can span centuries.
This entry covers how attribution works, who does it, what methods they use, how workshop practice complicates the question, and how machine learning is changing attribution research in 2025 and 2026.
How Attribution Works
Attribution rests on three pillars: documentary evidence, technical analysis, and connoisseurship. Documentary evidence includes contracts, inventories, letters, and early biographical accounts. If a 17th-century inventory lists a painting as "by Rembrandt," that is a starting point, though early inventories used artist names loosely, often to indicate workshop origin rather than the master's own hand. Technical analysis examines the physical materials and working methods: pigment types, canvas weave, priming layers, underdrawing, and pentimenti. If the materials are anachronistic, the attribution is wrong. If the working methods match the artist's known practice, the attribution is strengthened.
Connoisseurship is the visual judgment of an expert. The connoisseur has studied hundreds or thousands of works by the artist and can recognize the characteristic features of their hand: the way they draw ears, the pressure of their brush, the structure of their compositions. Connoisseurship is the oldest attribution method and remains the most common, but it is also the most subjective. Two equally qualified connoisseurs can disagree, and they frequently do.
The Morelli Method and Its Legacy
The modern science of attribution begins with Giovanni Morelli (1816-1891). Morelli was an Italian politician and art collector who had studied medicine before turning to art history. He applied a medical diagnostic approach to attribution, arguing that artists reveal their identity in the rendering of incidental details: the shape of ears, fingers, and toes, the folds of drapery, the treatment of hair. These details, Morelli claimed, were executed semi-consciously and were therefore harder to fake than major compositional elements. His method, published under the pseudonym Ivan Lermolieff in the 1870s and 1880s, was controversial but influential. Bernard Berenson in the United States and Roger Fry in Britain placed connoisseurship at the center of their practice, and the Morelli method shaped attribution scholarship for generations.
As art historian Amanda Wasielewski argued in her 2025 book "Computational Formalism: Art History and Machine Learning," reviewed in Leonardo journal, connoisseurship is "a type of formalism that serves a singular purpose: identification and authentication of an artwork." The review noted that computer-assisted connoisseurship has revived the category of style as a metric for sorting large digital art collections, despite the fraught history of the concept. Read the review at Leonardo.
The Workshop Problem
Attribution is straightforward when an artist works alone. It becomes genuinely difficult when the artist runs a workshop. From the Renaissance through the Baroque, most major painters operated large workshops where assistants and pupils executed substantial portions of commissions based on the master's designs. The master might paint the faces and hands, the most expressive passages, while assistants filled in backgrounds, drapery, and secondary figures. Who is the author of such a painting?
The case of Peter Paul Rubens (1577-1640) illustrates the problem. Rubens ran the most productive workshop in 17th-century Europe. His assistants included Anthony van Dyck and Jacob Jordaens, both major painters in their own right. A painting from Rubens's studio might contain passages by Rubens, passages by van Dyck, and passages by lesser assistants, all based on Rubens's design. A 2025 paper published on arXiv examined Rubens attribution using deep learning, noting that "the question of distinguishing between the master's hand and that of his assistants became one of the central tasks of Rubens studies, shaping generations of art historians and continuing to structure debates up to the present." The paper argued that machine learning is "particularly promising in cases involving recurring stylistic motifs and features of painterly technique on a micro-level. These features are imperceptible to the human eye but detectable through computational methods." Read the paper at arXiv.
A 2025 study published in "Kunstgeschichte" journal demonstrated the interplay of traditional connoisseurship and AI in a case involving a painting attributed to Van Dyck (1599-1641). The study combined art historical expertise with the AI authentication system developed by Art Recognition. The researchers noted that "in the 1630s, all works that came from Van Dyck's studio were at least partly produced collaboratively. Repetitions of his own paintings were created under Van Dyck's supervision in his workshop and were given a personal final touch, which ensured a seamless transition between copies and variants." Many paintings sold as Van Dyck's during his lifetime had never been touched by the master himself. The study's authors argued that AI should be trained by art historians and conservators who have spent decades studying the artist, not deployed as a standalone tool. Read the study at Kunstgeschichte.
Giovanni Morelli (1816-1891)
Morelli developed the first systematic method for attributing Italian Renaissance paintings based on the analysis of incidental anatomical details. His approach, rooted in his medical training, treated attribution as diagnosis: the artist is "given away by details of eyes, ears and knees, just as a criminal might be spotted by a fingerprint," as the archaeologist Michael Shanks put it. Morelli's method was controversial in his lifetime but became the foundation of modern connoisseurship.
Bernard Berenson (1865-1959)
Berenson was the most influential connoisseur of the early 20th century. Operating from Florence, he attributed thousands of Italian Renaissance paintings for collectors and dealers, including Isabella Stewart Gardner and Henry Clay Frick. His attributions carried enormous financial weight. Berenson was criticized for taking commissions from dealers whose paintings he authenticated, a conflict of interest that undermined his credibility in some cases.
Ludwig Burchard (1884-1960)
Burchard was the pioneering scholar of Rubens's oeuvre. His research formed the basis of the Corpus Rubenianum Ludwig Burchard, the ongoing project to publish a definitive catalogue of Rubens's work. The Corpus, organized into more than 30 parts, has been published progressively since the 1960s and remains the authoritative reference for Rubens attribution. Each part is written by a specialist who reexamines the attributions in light of current scholarship.
How Attribution Changes Over Time
Attributions are not permanent. A painting attributed to Rembrandt in 1900 may be reattributed to a workshop assistant in 2026, and vice versa. The Rembrandt Research Project, founded in 1968 by Ernst van de Wetering and others, systematically reexamined every painting attributed to Rembrandt. Over four decades, the project downgraded dozens of paintings from "Rembrandt" to "studio of Rembrandt" or "follower of Rembrandt," and in some cases upgraded paintings that had previously been dismissed. Van de Wetering later argued that the project's early downgrades had been too aggressive and that some workshop paintings deserved to be reconsidered as autograph.
The Van Eyck dispute of 2025 shows how new technologies can reopen settled questions. Art Recognition, a Swiss AI company, claimed that two versions of "Saint Francis of Assisi Receiving the Stigmata" attributed to Van Eyck were misattributed. The Van Eyck expert Maximiliaan Martens of Ghent University disputed the claim, arguing that only about 25 paintings are attributed to Van Eyck, an insufficient dataset for AI training. Martens called for peer-reviewed publications and collaboration with the scholarly community. Read the report at Artnet News.
Attribution is closely connected to authenticity, which is the verification that a work is genuinely by the attributed artist. It intersects with provenance, the ownership history that supports attribution claims. The conservator provides the technical analysis that can confirm or challenge an attribution. The atelier system is the workshop structure that makes attribution so difficult for Old Masters. The Old Master market depends on attribution for value. For more on how the art market handles questions of authorship, read our post on understanding art auctions or our guide on why provenance matters.
See Attribution Debates in Action
Attribution disputes are usually fought in scholarly journals, auction house back rooms, and courtrooms, but some play out in public. The National Gallery in London has reattributed several paintings in its collection over the past decades, and the changes are visible on the gallery labels. The Rubens Corpus Rubenianum project publishes its attributions in open-access volumes. The Rembrandt Research Project's corpus is available online, showing how attributions have shifted over time.
For more on how scholars determine who made a work of art, read our entries on authenticity and provenance, or explore our post on how to tell a painting's age without looking at the label to learn more about what materials and techniques reveal about a work's origins.
