BRENNEN

Industries / Retail, hospitality and personal services / Retail and e-commerce / Auction houses

Authenticity and ownership history

This is checking that a lot is what the catalogue says it is and that the seller has the right to sell it: that it is not a fake, not stolen, not looted from a person or a country, and not exported against the law.

6

Milestones achieved, from 1970 to 2016.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1970

    A treaty against illicit trade

    UNESCO's General Conference adopted the Convention on the Means of Prohibiting and Preventing the Illicit Import, Export and Transfer of Ownership of Cultural Property on 14 November 1970, giving its states parties a shared framework to stop trafficking and to return cultural property. The date still matters to sellers: EU import rules treat goods exported from a country before the 1970 Convention as a separate case.

    unesco.orgtaxation-customs.ec.europa.eu

    Credit

    UNESCO

  2. 1990

    A register of stolen art

    The Art Loss Register was set up in London in 1990 by shareholders from the insurance industry and the art market, growing out of a theft archive that the International Foundation for Art Research (IFAR) in New York had kept since 1976. It lists more than 700,000 lost, stolen and looted items and carries out more than 400,000 checks a year on items coming to market, for subscribing auction houses and art fairs, police, museums, dealers and others, by its own count.

    artloss.com

    Credit

    Art Loss Register; IFAR

  3. 1997

    Antiquities sales leave London

    After the journalist Peter Watson alleged in his book Sotheby's: The Inside Story that staff had helped take objects out of Italy, India and Cambodia against those countries' export laws, Sotheby's moved its regular antiquities and Indian art sales from London to New York, citing "issues of patrimony and heritage", and the heads of both departments resigned.

    LessonMoving a sale did not deal with objects that had no record of where they were found; the archaeologist Ricardo Elia told the magazine the scandals would end only when houses "stop dealing in unprovenienced antiquities".

    archive.archaeology.org

    Credit

    Peter Watson; Sotheby's; reported by Archaeology magazine

  4. 1998

    Principles for Nazi-looted art

    At the Washington Conference on Holocaust Era Assets, on 3 December 1998, 44 nations endorsed 11 non-binding principles: art confiscated by the Nazis and never returned should be identified, archives should be opened to researchers, unavoidable gaps in an object's ownership history should be taken into account, and a "just and fair solution" should be sought with pre-war owners or their heirs.

    state.govgetty.edu

    Credit

    US Department of State; the participating governments

  5. 2011

    A fake label exposes forgers

    Trasteco, a company based in Malta, bought Red Picture with Horses, sold as a 1914 painting by Heinrich Campendonk, for €2.4 million at the Lempertz auction house in Cologne in late 2006. The expert Ralph Jentsch judged its "Flechtheim Gallery" label a fake, and tests found titanium white, a pigment that did not exist in Campendonk's day. In June 2011 The National reported that prosecutors in Cologne had charged Wolfgang Beltracchi and three others over 14 fake paintings, with losses estimated at €34.1 million; the works had been sold as coming from the collection of a relative who, it turned out, never collected art.

    LessonAn invented ownership history held up for years because the Nazi period explained why works had gone missing.

    thenationalnews.com

    Credit

    Trasteco; Ralph Jentsch; Cologne prosecutors; reported by The National

  6. 2016

    Lab tests unmask a "Hals"

    Sotheby's had sold a portrait signed by Frans Hals privately in 2011 to the collector Richard Hedreen, for $10 million by The Art Newspaper's 2016 report and $11.2 million by its 2019 report; the Louvre had tried to buy it. In 2016 Orion Analytical, the lab of the scientist James Martin, found synthetic materials first produced in the 20th century in its paint layers. Sotheby's rescinded the sale and refunded the buyer in full, bought Orion in December 2016 to found an in-house scientific research department, and in 2019 settled its claim against the consignor, the dealer Mark Weiss, for $4.2 million.

    LessonPaint chemistry caught a forgery that experts and a national museum had accepted.

    theartnewspaper.comtheartnewspaper.comtheartnewspaper.com

    Credit

    James Martin, Orion Analytical; Sotheby's

Watch for

  1. AI attributions of paintings

    Art Recognition, a Swiss company, announced on 7 May 2026 that a Modigliani portrait, Jeune Femme au Col de Fourrure (1915), was "the first AI-authenticated artwork to be accepted into an official catalogue raisonné", the complete listing of an artist's works, and gave it a 78.38% probability of being genuine. The scholar who compiled that catalogue raisonné, Marc Restellini, said he "never used AI for my catalogue", and the company republished an edited release on 22 May.

    Achieved whenAn independent blind test on works whose status is settled, including known forgeries, publishes the method's error rates.

    theartjournal.com

    Claimed by

    Art Recognition, May 2026

  2. Provenance reports in minutes

    ArtCheck says it searches 24 museum, authority and loss-register databases for an object's ownership history and returns a report "in minutes", including a Nazi-era risk assessment. It calls each report "research triage" and "not a verdict", and gives no accuracy figure.

    Achieved whenAn outside test scores its reports against cases whose history researchers have already settled.

    artcheck.work

    Claimed by

    ArtCheck, on its website, read November 2026

  3. Drawings judged by five measures

    Researchers at the University of Bradford, the Independent Media Institute and Stanford University report a system that uses five visual features, such as texture and contrast, to judge 900 authenticated drawings by ten artists, with 89.8% overall accuracy: it accepted genuine works 83% of the time and rejected impostors more than 90% of the time. They say it can learn an artist from as few as 20 drawings, and the release says the work was published in PLOS ONE.

    Achieved whenOthers test it on drawings it has not seen, including documented forgeries.

    bradford.ac.uk

    Claimed by

    Hassan Ugail and colleagues, June 2026

Open

  1. 01

    Blind test of AI authentication

    A sealed set of images of 100 paintings and drawings whose status is settled, mixing works with secure attributions, documented forgeries and old copies, sent to every AI authentication service that will take them, with the answer key kept from the services until all results are in.

    The brief
    Credit · unclaimed
  2. 02

    Score provenance-screening tools

    A set of 300 ownership histories, written as catalogues print them, in which researchers have already settled whether the object was seized or sold under Nazi persecution, plus clean histories of similar length, and a published score for each screening tool and for plain AI models on the set.

    The brief
    Credit · unclaimed

Open 01 · the brief

Blind test of AI authentication

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A sealed set of images of 100 paintings and drawings whose status is settled, mixing works with secure attributions, documented forgeries and old copies, sent to every AI authentication service that will take them, with the answer key kept from the services until all results are in.

Why it is still open

Specialists dispute Art Recognition's attributions, and the scholar it named said he had not used its AI. The search for this file found no independent test of any commercial AI authentication service on works with a known answer.

theartjournal.com

Built with what exists

Museum open-access images for secure works; published images of exposed forgeries, such as the Beltracchi fakes and the "Hals" portrait; the services' own submission routes; the answer key stored with a timestamp in a public repository before any result comes back.

  1. First step

    List 30 forgeries exposed in court cases or museum reports since 1990 for which good images are public, and note the licence of each image.

  2. Done looks like

    Each service's calls on all 100 works published as a table of right, wrong and declined, with forgeries wrongly accepted counted separately, and the answer key released after scoring.

Open 02 · the brief

Score provenance-screening tools

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A set of 300 ownership histories, written as catalogues print them, in which researchers have already settled whether the object was seized or sold under Nazi persecution, plus clean histories of similar length, and a published score for each screening tool and for plain AI models on the set.

Why it is still open

Tools now screen catalogues for gaps between 1933 and 1945 and for names linked to looting, but ArtCheck calls its own output "triage" and "not a verdict" and gives no accuracy figure, and the search for this file found no published accuracy test of any such tool.

artcheck.work

Built with what exists

Published decisions of national restitution panels and museum restitution reports; the German Lost Art database and the ERR Project database of objects seized in occupied France; the Getty Provenance Index; Claude Fable 5.2 as a baseline (see the Note "Three working devices") and Gemini Decision 1, which returns probabilities for fixed-answer questions (see the Note "A thousand questions, one pass").

  1. First step

    Collect 50 published restitution decisions that quote the object's ownership history in full.

  2. Done looks like

    For each tool, the share of known looting cases it flags and the share of clean histories it wrongly flags, on all 300 cases, with two researchers agreeing on every label before scoring.