Research

The science behind Know Your Object.

3 articles · 9 references

01 · Regulation & markets

Why art needs Know Your Object

Banks know their customers. Until now, nobody had a standard way to know the object.

Rembrandt’s The Night Watch: a militia company steps out of the dark, led by a captain in black and a lieutenant in pale yellow.
Rembrandt van Rijn, The Night Watch, 1642 · Rijksmuseum, Amsterdam · Public domain
28 Jun 2025
EU import rules for cultural goods applyReg. (EU) 2019/880
€10,000
AML threshold per transaction for art traders, from 10 July 2027Reg. (EU) 2024/1624
$33.9–40bn
Estimated art-secured lending marketDeloitte, Art & Finance Report 2025
$59.6bn
Global art sales in 2025Art Basel & UBS, 2026

KYC has no counterpart for art

Before a bank moves money, it checks who its customer is. Know Your Customer (KYC) is routine, documented and auditable. Artworks never had an equivalent. A painting can be sold, insured, lent against and shipped across borders, and each party checks it in its own way, often on paper and from scratch. There is no shared record that says: this is the object, this is its history, and this is who checked it.

The rules are catching up

Since 28 June 2025, EU Regulation 2019/880 governs the import of cultural goods into the Union. Importers of cultural goods older than 200 years with a customs value above €18,000 must file an importer statement. Archaeological objects older than 250 years need an import licence.⁠1

Anti-money-laundering rules follow. Under EU Regulation 2024/1624, traders in works of art take on AML obligations from 10 July 2027 for transactions of €10,000 or more.⁠2 For dealers and auction houses, checks that used to be good practice become a compliance task that must be documented.

Art is already collateral

Art is not only collected. It is financed. Deloitte estimates the art-secured lending market at US$33.9–40 billion.⁠3 Global art sales reached US$59.6 billion in 2025.⁠4 Every loan, policy and sale in that market rests on the same question: is this the object everyone thinks it is?

What KYO adds

KYO applies the logic of KYC to the object. One report combines an optical fingerprint, provenance and watchlist checks, so that a bank, an insurer and an auction house can rely on the same verified identity instead of repeating the work. The next two articles explain the science behind the first two layers.

02 · Optical fingerprint

Craquelure: the fingerprint in the paint

Old paintings carry a pattern nobody painted on purpose. Researchers are teaching machines to read it.

Detail of Vermeer’s The Milkmaid: bread, a jug and a blue cloth on a table, with a fine network of cracks across the paint.
Johannes Vermeer, The Milkmaid, c. 1660 (detail) · fine cracks run across bread and cloth · Rijksmuseum · Public domain

A network nobody painted

As a painting ages, its paint and ground layers crack. The result is craquelure: a fine network of lines spread across the surface. The artist did not design it. It grows out of the materials, the support and the life of the object. That is what makes it interesting for identification. The network belongs to one specific painting, which is why researchers treat it as a fingerprint.

Teaching a model to read cracks

Chirosca and colleagues, writing in Applied Sciences in 2025, use an autoencoder: a neural network that learns to compress an image into a compact code and rebuild it from that code. Applied to craquelure, it turns a crack network into a signature that can be stored and compared. Two images of the same painting should lead to matching signatures. A different painting should not.⁠5

The abstract reports no accuracy figures. We cite the paper for its method and quote none.

The canvas has a signature too

Beneath the paint, the canvas carries its own pattern. Murillo-Fuentes, Olmos and Alba-Carcelen trained a Siamese deep-learning model to compare canvas fabrics and applied it to paintings from the Museo Nacional del Prado. Their approach compares weaves even when thread densities look alike.⁠6 For a verifier, that is a second physical signature, independent of the paint surface.

Preprint, submitted 25 June 2025. Not yet peer-reviewed.

Why it matters for verification

A fingerprint turns “this looks like the painting” into “this matches the record”. Captured once, it can be checked again at every sale, loan or shipment. That is the idea behind the optical fingerprint in a KYO report: identity anchored in the physical object itself, not only in its paperwork. Paper can be copied or lost. The pattern travels with the painting.

03 · Provenance & materials

Reading the archive with AI

An artwork’s history lives in catalogues, records and letters. Most of it was never meant to be searched.

Detail of Rembrandt’s The Jewish Bride: two hands meet on a red gown, painted in thick, textured strokes.
Rembrandt van Rijn, Isaac and Rebecca (‘The Jewish Bride’), c. 1665–1669 (detail) · Rijksmuseum · Public domain

Provenance is a paper trail

Provenance is the documented chain of ownership of an object: who held it, when, and how it changed hands. The evidence sits in auction catalogues, dealer records, inventories and correspondence, spread across many archives and often handwritten. Checking it by hand is slow. That makes provenance one of the bottlenecks of any verification.

Handwriting becomes data

Handwritten text recognition changes this. Muehlberger, Seaward, Terras and colleagues describe how the Transkribus platform turns archival manuscripts into searchable text, and how that transforms research in the archives.⁠7 Once a record is text, it can be searched, linked and compared.

Researchers now apply the same idea to provenance directly. A team at Santa Clara University is building an assistant that lets users query about 30,000 Jeu de Paume records of looted art in plain language, with AI explaining each match.⁠8

Early-stage research tool, limited to one database. Reported as a university news story, not a peer-reviewed paper.

Reading the materials, too

AI can also read the object itself. Macro X-ray fluorescence (MA-XRF) scanning maps the chemical elements across a painting’s surface and produces large spectral datasets. Preisler and colleagues show in Science Advances (2024) that deep learning analyses these spectra more accurately than classic methods and strips out artifacts. They demonstrate it on a painting by Raphael.⁠9

The machine proposes, a human signs

None of this replaces expertise. Models surface candidates: a matching record, an unusual element map, a gap in the chain of ownership. A human expert reviews the evidence, decides and signs off. KYO is built around that split. AI does the reading. People take responsibility for the result, and every step stays traceable from the source record to the final signature.

Sources

References

  1. Regulation (EU) 2019/880 of the European Parliament and of the Council on the introduction and the import of cultural goods. EUR-Lex.

    eur-lex · 2019/880 (opens in a new tab)
  2. Regulation (EU) 2024/1624 of the European Parliament and of the Council on the prevention of the use of the financial system for the purposes of money laundering or terrorist financing. EUR-Lex.

    eur-lex · 2024/1624 (opens in a new tab)
  3. Deloitte Private & ArtTactic. Art & Finance Report 2025. Deloitte.

    deloitte.com (opens in a new tab)
  4. Art Basel & UBS. The Art Basel and UBS Global Art Market Report 2026. Art Basel.

    artbasel.com (opens in a new tab)
  5. Chirosca, Radvan, Pop, Chirosca (2025). Autoencoder Application for Artwork Authentication Fingerprinting Using the Craquelure Network. Applied Sciences 15(16), 9014.

    doi:10.3390/app15169014 (opens in a new tab)
  6. Murillo-Fuentes, Olmos, Alba-Carcelen (2025). Forensic Study of Paintings Through the Comparison of Fabrics. arXiv preprint, not peer-reviewed.

    arXiv:2506.20272 (opens in a new tab)
  7. Muehlberger, Seaward, Terras et al. (2019). Transforming scholarship in the archives through handwritten text recognition: Transkribus as a case study. Journal of Documentation 75(5), 954–976.

    doi:10.1108/JD-07-2018-0114 (opens in a new tab)
  8. Santa Clara University (2026). AI Provenance Assistant. University news story, D. Lohse, 5 July 2026. news.scu.edu.

    news.scu.edu (opens in a new tab)
  9. Preisler, Andolina, Busacca, Caliri, Miliani, Romano (2024). Deep learning for enhanced spectral analysis of MA-XRF datasets of paintings. Science Advances 10(39).

    doi:10.1126/sciadv.adp6234 (opens in a new tab)

From research to report.

Request access for auction houses, lenders and insurers.