Italian police used AI to recover a stolen painting hidden beneath overpaint

The woman in the painting held a scroll. The stolen saint investigators were looking for held a lamb. New BBC reporting describes how Italian police used AI to connect those apparently different images, recovering an altered 17th-century painting spotted at a Parma antiques fair in October 2024—two decades after its theft.

Officers visiting the fair photographed artworks and uploaded the images to an internal database. Within hours, their system flagged the portrait. The next day, officers seized the painting so art experts could examine it more closely.

Investigators identified it as Saint Agnese, stolen from the church of Virgo Potens in Sulmona on August 11, 2004. Its appearance had changed: the lamb had become a scroll, while the halo and palm branch had disappeared. The apparent overpainting turned a religious figure into a young woman without those identifying symbols.

A keyword search for a woman with a scroll would not have retrieved the stolen painting. That mismatch explains the value of the Stolen Works of Art Detection System, or Swoads: it compared visual features rather than relying on a description of what the woman was holding.

Lt Col Salvatore Rapicavoli, of the Carabinieri’s cultural heritage protection unit, told the BBC that the software matches similarities in colour, rough outlines and composition. It can also identify parts of paintings that have been cut up and rearranged. Once it retrieved the older image, the resemblance was readily apparent to investigators.

The team conceived Swoads in 2020 to tackle a laborious comparison job. Italy’s stolen-cultural-property database contains more than one million items, though only some records include photographs. Officers previously searched descriptions and scrolled through images, sometimes working from decades-old, black-and-white pictures with glare, obscured details or extreme contrast.

The software also continuously scans online catalogues and new uploads against the database. Rapicavoli credited it with helping recover more than 300 large paintings and statues, plus more than 165,000 smaller objects, worth over €203 million combined. Those are the team’s reported recovery totals, not an independent measure of the system’s accuracy.

You have to have a large team dedicated to nothing but chasing leads generated by the AI

Timothy Carpenter, former head of the FBI’s Art Crime Team, speaking to the BBC

Carina Popovici, co-founder of art-authentication company Art Recognition, drew a narrower boundary around the technology’s role. Such systems ask whether an image matches something reported stolen. They cannot flag an unrecorded theft or catch the thief; recovery still requires investigators and legal cooperation.

Carpenter identified another complication: an image match does not necessarily distinguish an original from a copy. Stolen-art databases can contain multiple versions of the same composition and do not require proof of authenticity before recording a work. A convincing resemblance therefore does not settle which version investigators have found.

Technical blind spots remain, too. Poor training images can impair recognition, and three-dimensional objects pose a particular problem: a vase photographed from the front may not match a sale listing showing its back. Rapicavoli’s team is working to improve that capability.

The works most likely to surface are not necessarily the most famous. Researcher Steven Frank told the BBC that high-profile stolen pieces may avoid ordinary sales channels altogether. Rapicavoli likewise said he did not expect Swoads to locate Caravaggio’s missing Nativity. The Saint Agnese case instead shows its usefulness when a disguised work re-enters a market police can search.