Attribution in 2026: Algorithm, Laboratory, Court

Two van Eycks, two numbers, three incompatible readings

On 9 February 2026 the Swiss company Art Recognition reported that neither version of Saint Francis of Assisi Receiving the Stigmata is by the hand of Jan van Eyck. Its model returned a negative result at 91 per cent for the panel in the Philadelphia Museum of Art and 86 per cent for the larger version in the Galleria Sabauda, part of the Musei Reali in Turin. Carina Popovici, the company's co-founder and director, told the Guardian that she had expected one of the two to come back positive, that both were negative, and that she did not think the museums would be pleased. The Philadelphia Museum of Art said only that it had not been involved in the analysis and was therefore unable to comment. The Turin museums did not respond to requests for comment.

The finding is not the interesting part. Three sets of people read the same two numbers three different ways. The press reported a claim that the pictures are fakes. The company's own formulation is narrower: not by the hand of Van Eyck. Two Van Eyck specialists took the numbers as support for something already on the table. Till-Holger Borchert, formerly director of the Bruges city museums, said the result supports earlier suggestions that both are studio paintings and mainly raises questions requiring further investigation. The art historian Noah Charney, on whose podcast the finding was discussed, drew out the consequence: All this could mean that there is a lost original somewhere that is much more Van Eyck's work than these copies.

Between a forgery, a workshop product and the shadow of a lost prototype there is no shared price, no shared label and no shared cataloguing decision. One percentage cannot carry that much freight.

The physical evidence points the other way, and the surface is a poor witness

The Philadelphia Museum of Art published the technical dossier on both pictures in 1997 as Jan van Eyck: Two Paintings of Saint Francis Receiving the Stigmata. The Philadelphia picture is oil on vellum laid on panel, 12.4 by 14.6 cm, John G. Johnson Collection, cat. 314; the Turin picture is oil on panel, 29.2 by 33.4 cm, cat. 187. Both are dated there to the 1430s, the 2026 coverage to about 1428-32.

Two findings in it matter now. Peter Klein dated the growth rings of the Philadelphia panel between 1225 and 1307 against the Baltic and Polish master chart, and matched its curve to the panels of Van Eyck's portraits of Giovanni Arnolfini and Baudouin de Lannoy in Berlin, the wood coming from the same tree. J. R. J. van Asperen de Boer's infrared reflectography found extensive Eyckian underdrawing in the Turin painting and only a few possible strokes in the Philadelphia one, which led Marigene H. Butler to conclude that Turin may well be one of the two Saint Francis pictures by Van Eyck's hand named in Anselm Adornes's will of 1470. The shared-tree evidence sits under Philadelphia, the underdrawing evidence under Turin, and the model scored Philadelphia the more negative. A panel from workshop stock establishes where the wood came from, not who painted the parchment glued to it.

Maximiliaan Martens of Ghent University rejects the exercise on grounds that go to what a surface can be asked. Training a model on the painter's brushstroke attempts the wrong thing: Even when one studies Van Eyck's paintings on a microscopic level, his brushstrokes are barely visible. That is one of the most prominent characteristics of his work! He adds that both pictures have been damaged and retouched across six centuries, that the surviving corpus is too small to train on, and that without a serious peer-reviewed article the claims are thoroughly untrustworthy. Popovici replied that Borchert helped build the dataset and a paper was in preparation.

Condition is not a footnote. When the Berlin Gemaeldegalerie withdrew The Man with the Golden Helmet from Rembrandt in 1986, the examination found the canvas cut down on all sides and large areas of the face painted over beneath an ochre varnish. Authorship had been judged on a surface substantially not the painter's.

The Rijksmuseum showed what a finished attribution argument looks like

Three weeks later, on 2 March 2026, the Rijksmuseum announced that Vision of Zacharias in the Temple, dated 1633, is by Rembrandt. The picture was excluded from the oeuvre in 1960, bought by a private individual in 1961 and lost to scholarship for sixty-five years until the current owner approached the museum. Artnet, citing Het Parool, reports that specialists after 1960 considered Jan Lievens or Salomon Koninck, a secondary chain wanting a check against the 1960 literature.

The two-year study used the techniques developed for Operation Night Watch. Macro-XRF scanning and visual inspection revealed compositional changes; the pigments were found in other Rembrandts of the period; the paint layer build-up matched his early work; the signature was found to be original; and dendrochronology of the panel confirmed the painted date of 1633. The museum's summary is deliberately cumulative: Materials analysis, stylistic and thematic similarities, alterations made by Rembrandt, and the overall quality of the painting all support the conclusion that this painting is a genuine work by Rembrandt van Rijn. The comparanda were named, among them Daniel and Cyrus Before the Idol Bel of 1633 in the J. Paul Getty Museum and Jeremiah Lamenting the Destruction of Jerusalem of 1630 in the Rijksmuseum. Jonathan Bikker and Petria Noble published the full account in the March 2026 Burlington Magazine, volume 168, number 1476, behind a subscription. The painting went on view on 4 March, on long-term loan from a private collector.

The contrast is procedural and it is the whole of the matter. One institution named its methods, its comparanda and its journal in the same breath as its conclusion. The other issued a number and left the paper for later.

In Italy nobody can go to court to be told what a picture is

The Turin Saint Francis hangs in an Italian state museum, which makes the Italian position on authenticity load-bearing. On 9 February 2025 the Corte di Cassazione, in decision number 3231, held that a standalone action seeking only a judicial declaration that a work is authentic is inadmissible. Authenticity may be assessed by a court only inside a wider dispute, tied to damages, breach of contract or some other economically relevant harm. Italian courts have also consistently treated the opinions of archives and foundations as expressions of intellectual and scholarly judgment protected by freedom of expression, with the consequence that a court cannot compel an archive to issue a certificate or to enter a work in a catalogue. This account comes from the Chambers Art and Cultural Property Law 2026 guide, Italy chapter, by Angela Saltarelli; the case number is given so the judgment can be read direct.

The practical effect is that a disappointed owner must reframe the question as money. Attribution stops being a fact to be established and becomes an element of loss to be proved. That is the ceiling under every attribution story of 2026: the gatekeeper's opinion is not reviewable on its merits, only its consequences are.

Documentation duties moved in the same period. Law number 40 of 17 March 2026 raised the threshold for permanent export of works of art from Italy to EUR 50,000, books and library material excepted at EUR 13,500. Italian dealers have carried customer due diligence and suspicious transaction reporting at or above EUR 10,000 since Legislative Decree 125/2019.

What it costs when they change their minds

The market consequence is legible in a single object. Rembrandt's Adoration of the Kings was catalogued as a Rembrandt by Hannema in 1955, by van Gelder in 1957 and by Benesch in 1957 and again in 1973. Bauch called it a Rembrandt pupil after 1630 in 1960. Sumowski recorded it as a studio work of about 1630; the Sotheby's catalogue cites his 1983 volume at entry 1979, while Artnet dates the attribution to 1979, apparently reading the entry number as a year. It failed to sell at Sotheby's on 27 March 1963 as a Rembrandt at GBP 3,800, sold at Christie's Amsterdam on 3 December 1985 as circle of Rembrandt, and sold again in a Christie's Amsterdam online sale closing 6 October 2021, still as circle of Rembrandt.

It returned to Sotheby's London on 6 December 2023 as lot 11, catalogued as Rembrandt, estimated at GBP 10,000,000 to 15,000,000, flagged a premium lot, guaranteed property and subject to irrevocable bids. The lot page does not publish the result without an account. Coverage reports GBP 10.9 million, or USD 13.7 million, without stating whether that is the hammer or the premium-inclusive total, and the dollar figure is a conversion at an unstated date. On a guaranteed lot with an irrevocable bid the reported total is a weak reading of demand.

Our own records hold the same mechanism without the upside. The Camden attic Portrait of a Girl, hammer USD 1,175,000 on 24 August 2024, was published as a Rembrandt by Bode and hung as one in the Mauritshuis, and sold with no published argument for its demotion. The Getty kouros, labelled Greek, about 530 B.C., or modern forgery, faced nineteen invited specialists in Athens in May 1992 who could not agree. The Met's Crucifixion and Last Judgment were argued between six candidate hands for over a century.

The only rulebook is voluntary and its signatories are its subjects

There is a governance document. In November 2025 the Center for Art Law published Guidelines for AI and Art Authentication, with a longer framework paper produced with Art Recognition. The page states its four principles twice and not identically: first as transparency, accountability, scientific rigor and human-AI collaboration, then as transparency, scientific rigor, ethical responsibility and human-AI collaboration. It also carries the line that settles its legal weight, that the resource is for educational purposes only and does not offer legal advice.

The signatures are the point. Popovici endorsed it herself, writing that the foundational principles must guide both the development of robust AI authentication models and the responsible, well-informed integration of AI into art authentication practices. Irina Tarsis of the Center for Art Law framed the tool as an addition to an existing structure: The so-called three legged chair of art authentication that relies on scientific analysis, provenance research and connoisseurship now has an extra point of support - AI-systems. Charney, who endorsed the guidelines as best practice, is also the host on whose programme the van Eyck result was aired, and Borchert both helped assemble the dataset and commented publicly on what it produced. The pool qualified to judge Van Eyck is small enough that building the instrument and reviewing its output fall to the same few names.

Three months after those principles appeared, the specialist objection to the van Eyck claim was that no peer-reviewed article existed. That is not an outsider's complaint about a new technique. It is the framework's own transparency requirement, unmet by one of its authors, with no mechanism anywhere that could have required otherwise.