Industries / Media, arts and recreation / Art galleries
Authenticity and ownership history
This is the work of checking that a work is by the artist named, that the seller has the right to sell it, and that its record of past owners, called its provenance, holds up.
6
Milestones achieved, from 1990 to 2022.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1990
A register of stolen art
The Art Loss Register was set up in London in 1990, growing out of a theft archive that the International Foundation for Art Research began in 1976. It now lists more than 700,000 lost, stolen and looted items and runs more than 400,000 checks a year for auction houses, art fairs and dealers.
Credit
the Art Loss Register; the International Foundation for Art Research
2011
Authenticators step back
In October 2011 the Andy Warhol Foundation said it would close its authentication board early in 2012. The board had faced legal disputes from owners of works it would not accept as Warhol's.
LessonWhen saying "not genuine" draws lawsuits, experts stop giving opinions, and buyers lose a check.
Credit
the Andy Warhol Foundation for the Visual Arts
2013
Fakes sold through Knoedler
On 16 September 2013 the dealer Glafira Rosales pleaded guilty to selling more than 60 fakes of works by Rothko, Pollock, de Kooning and Motherwell to the now-closed Knoedler gallery and another gallery, for about $33.2 million. Collectors had paid more than $80 million for them, and she agreed to forfeit $33.2 million and pay $81 million in restitution.
LessonAn old gallery name was no substitute for checking where works came from.
Credit
the US Attorney's Office for the Southern District of New York
2016
A dealer's hoard examined
A German task force reported on 14 January 2016 on almost 1,500 works held by Cornelius Gurlitt, son of Hildebrand Gurlitt, a dealer who worked for the Nazis. Research pointed to 162 works as Nazi loot, but only five were definitely identified as taken from Jewish families, and provenance was settled for 13.
LessonWhen a dealer's business was "marked by discretion, secrecy and deception", two years of research left most works unresolved.
Credit
the Gurlitt task force
2022
Dealer stock books as data
The Getty Provenance Index publishes the stock books of M. Knoedler & Co. (more than 40,000 records, 1872–1970) and Goupil & Cie (more than 43,000 records, 1846–1919) as downloadable files in the public domain under CC0. The current tagged release, v3, dates from 1 June 2022.
Credit
the Getty Research Institute
2022
Hidden shares in artworks
In May 2022 the dealer Inigo Philbrick was sentenced to seven years in prison for wire fraud. He had borrowed against and resold fractional shares in costly contemporary works, using concealed ownership interests, fake documents and an invented collector, and he agreed to pay $86 million and forfeit two paintings.
LessonBuyers and lenders had no shared record showing who already held a stake in a work.
Credit
the US Attorney's Office for the Southern District of New York
Watch for
AI backs a disputed Caravaggio
Art Recognition, a Swiss company, says its model gives an almost 86% chance that the version of The Lute Player formerly at Badminton House is by Caravaggio. The firm says it can judge authenticity from a photograph, using models trained on accepted works and on copies and imitations. Experts quoted by CNN called this "quite problematic" and cited "a lack of transparency", since outsiders cannot inspect the models or data.
Achieved whenan independent technical study of the painting, or a blind test of the model run by outsiders, supports the attribution.
Claimed by
Art Recognition, September 2025
Provenance text parsed at scale
A parser built at the University of Basel turned the free-text ownership histories of 48,535 Rijksmuseum works into 101,115 structured events. Fixed rules handle about 97% of events and Claude, an AI model, the rest. The project says its AI-made additions "have not been reviewed or endorsed by the Rijksmuseum" and publishes no accuracy figures.
Achieved whensomeone outside the project publishes precision and recall on a hand-checked sample, or the Rijksmuseum adopts the data.
Claimed by
Arno Bosse, RISE, University of Basel, 2026
Open
01
A blind attribution test
A sealed test set of photographs of paintings whose status is settled: accepted works, workshop and later copies, and known forgeries, for 20 artists. It is sent to commercial attribution services and open models, with the answer key fixed in advance and published only after the results come back.
The briefCredit · unclaimed02
Provenance gaps on the market
A count of the works offered for resale by galleries and fairs whose published provenance has a gap between 1933 and 1945, the Nazi era. The provenance text is parsed with an existing open parser, and only totals by price band and region are published, not accusations against any seller.
The briefCredit · unclaimed
Open 01 · the brief
A blind attribution test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A sealed test set of photographs of paintings whose status is settled: accepted works, workshop and later copies, and known forgeries, for 20 artists. It is sent to commercial attribution services and open models, with the answer key fixed in advance and published only after the results come back.
Why it is still open
Outside researchers cannot check commercial models without access to "commercially sensitive models and datasets" (CNN). The published studies test research models instead. One is a 2024 paper on van Gogh forgeries, including fakes by Otto Wacker, co-written by Art Recognition staff. Another is a 2025 test of six general vision-language models, which found them unreliable at naming artists. Neither tested a commercial service blind.
Built with what exists
Open-access collection images from museums such as the Rijksmuseum and the Metropolitan Museum of Art; Meta's open DINOv2 image model as a baseline; a small vision model from Liquid AI's LFM3 family tuned on one graphics chip (see the Note "One GPU to fine-tune"); a hash of the answer key posted to a public git repository before any results arrive.
First step
Build the set for one artist with many catalogued copies from a museum's open data, and post the hash of its answer key.
Done looks like
At least 30 accepted works and 30 other works for each of 20 artists, and each system's accuracy reported with confidence intervals. At least two commercial services take part, or their refusal is reported.
Open 02 · the brief
Provenance gaps on the market
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A count of the works offered for resale by galleries and fairs whose published provenance has a gap between 1933 and 1945, the Nazi era. The provenance text is parsed with an existing open parser, and only totals by price band and region are published, not accusations against any seller.
Why it is still open
The Basel parser works on one museum's own records, and its AI-made output is unreviewed. This research found no count of provenance gaps for works on the market.
Built with what exists
The rijksmuseum-mcp+ parser, released under the MIT licence; Browser Use 1.0 to collect listings (see the Note "Deleting 80% to start over"); Muse Ember for extraction (see the Note "The jobs agents repeat"); a second model pass to catch misses (see the Note "Seven more flagged").
First step
Collect provenance text from 100 resale listings on gallery websites, run the parser on it unchanged and record where it fails.
Done looks like
2,000 listings parsed, with 200 checked by hand showing at least 95% of ownership events correctly extracted, and the share with a 1933–1945 gap published by price band.