# Attribution in 2026: Algorithm, Laboratory, Court > In February an algorithm scored two van Eycks negative; in March the Rijksmuseum published a Rembrandt. One side showed its working. Italian law will adjudicate neither. > The Study - 2026-08-02 > Canonical HTML: https://iconotheca.com/study/attribution-2026-algorithm-laboratory-court/ ## Key points - Art Recognition's model returned 91 per cent negative on the Philadelphia Saint Francis and 86 per cent on the Turin version, but the published technical record puts the strongest physical evidence under the Philadelphia picture, not the Turin one. - A 91 per cent negative score is not a 91 per cent probability of forgery: the same number was reported as a fake, stated by the company as not by his hand, and read by two Van Eyck specialists as evidence of studio production. - The Rijksmuseum's reattribution of Vision of Zacharias in the Temple on 2 March 2026 named its methods, its comparanda and its publication venue in the same announcement. The van Eyck claim reached the press before the peer-reviewed paper. - The Italian Supreme Court held on 9 February 2025 that nobody may sue merely to have a work declared authentic, and Italian courts will not compel an archive to certify one, so the Turin picture's status cannot be adjudicated in the abstract where it hangs. ## Two van Eycks, two numbers, three incompatible readings ## The physical evidence points the other way, and the surface is a poor witness ## The Rijksmuseum showed what a finished attribution argument looks like ## In Italy nobody can go to court to be told what a picture is ## What it costs when they change their minds ## The only rulebook is voluntary and its signatories are its subjects The two announcements of early 2026 are a controlled experiment in how attribution gets decided, and the Rijksmuseum won it on procedure rather than on prestige. It published its methods, its comparanda and its journal at the moment it published its conclusion, so the claim can be checked, argued with and eventually overturned by anyone willing to do the work. A percentage released ahead of its paper cannot be checked, and the Italian ruling of February 2025 means no court will check it either. The instruction for a file is narrow and firm: record who made the claim, on what material, with what published argument, and on what date. Attribution is a state of the record about an object rather than a property of the object, and the record is the only part anybody can be held to. ## Questions **Does a 91 per cent negative result mean there is a 91 per cent chance the picture is a forgery?** No, and the distinction is the one that gets lost in reporting. The company's stated conclusion is that the work is not by the hand of the artist, which in a fifteenth-century workshop context is consistent with a studio product made under the master's direction. Both Borchert and Charney read the two Saint Francis scores that way rather than as evidence of fraud. **Does the same-tree dendrochronology settle the Philadelphia picture in Van Eyck's favour?** It does not. Klein's result ties the wooden support to the same tree as two accepted Van Eyck portraits in Berlin, which places the panel in the workshop's material supply. The Philadelphia image is painted on vellum laid onto that panel, so the wood evidence speaks to the support rather than to the hand that made the picture. It is strong circumstantial evidence and it is not an attribution. **If an authentication body refuses my work, can I litigate the attribution itself?** Under Italian law as clarified on 9 February 2025, no. A standalone declaratory action on authenticity is inadmissible, and a court will not order an archive to certify a work or to include it in a catalogue. The question can only be reached inside a claim for damages, breach of contract or comparable economic harm, which means the pleading has to be about loss rather than about the picture. **What should be in a file before an AI attribution result is treated as evidence?** The composition of the training set and who assembled it, the treatment history and current condition of the object photographed, the imaging conditions, the direction and magnitude of the reported score with its stated meaning, and a peer-reviewed account of the method. On the van Eyck claim of February 2026 the last of these was still in preparation, which is why a specialist could dismiss the result without disputing the technique in principle. ## Sources - [1] Artnet News, Jo Lawson-Tancred, 'Van Eyck Attribution Dispute Pits Art Historians Against A.I. Firm', 9 February 2026, updated 10 February and edited 18 February 2026. - [2] VRT NWS, Colin Clapson, 'AI challenges authenticity of works by Flemish master Jan van Eyck', dated on the article 14 February 2026; the URL path carries 11 February 2026. - [3] Philadelphia Museum of Art, 'Jan van Eyck: Two Paintings of Saint Francis Receiving the Stigmata', 1997, ISBN 0-87633-115-0, ed. Jane Watkins; full text read from the Internet Archive scan. Source of the plate data, Peter Klein's dendrochronology and van Asperen de Boer's reflectography. - [4] Rijksmuseum press release, 'Rijksmuseum researchers discover new painting by Rembrandt van Rijn', 2 March 2026, 10:00. - [5] The Burlington Magazine, Jonathan Bikker and Petria Noble, 'Rembrandt's Vision of Zacharias in the Temple', March 2026, vol. 168, no. 1476. Title, authors and issue verified; the text is behind a subscription and was not read. - [6] Artnet News, Jo Lawson-Tancred, 'Rediscovered Rembrandt Confirmed After Decades of Doubt', 2 March 2026. Source of the Het Parool report on the Lievens and Koninck alternatives, which is tertiary here and should be replaced with the 1960 literature. - [7] Chambers and Partners, 'Art & Cultural Property Law 2026', Italy, Trends and Developments, Angela Saltarelli, Jacobacci Avvocati, last updated 14 April 2026. A practitioner guide, not the judgment; Corte di Cassazione 9 February 2025 no. 3231 and Law no. 40 of 17 March 2026 should be read in the primary text. - [8] Center for Art Law, 'Guidelines for AI and Art Authentication', last updated 14 November 2025. Source of the four principles as stated twice on the page, the educational-purposes disclaimer, and the Popovici, Tarsis and Charney statements. - [9] Center for Art Law with Art Recognition, 'Framework for Responsible Use of AI in Art Authentication', November 2025. The PDF was downloaded but its embedded text could not be extracted, so no claim here rests on its interior wording. - [10] Sotheby's, lot 11, Rembrandt Harmensz. van Rijn, 'The Adoration of the Kings', Old Master & 19th Century Paintings Evening Auction, London, 6 December 2023. Source of the estimate, the premium-lot and guarantee flags, and the printed attribution history from Hannema 1955 to Sumowski 1983. The result is not shown without an account. - [11] Artnet News, Adam Schrader, 'A Newly Attributed Rembrandt Once Valued at $10,000 Sold for Almost $14 Million', 13 December 2023. Reports GBP 10.9 million and USD 13.7 million without distinguishing hammer from premium, and dates Sumowski's attribution to 1979.