During a routine visit to the annual antiquity fair in Parma in October 2024, Italian police officers walked between the stands snapping photographs of the artworks. It was not art appreciation. The images went straight into an internal database.
A couple of hours later the system flagged one of the snapshots: a painting of a young, fair-skinned woman in aquamarine and red, holding up a half-unravelled parchment scroll.
“There was something off,” says Lt Col Salvatore Rapicavoli, of the Italian police unit for the Protection of Cultural Heritage. Officers returned the next day and seized the painting for their art experts.
The disguise, and why it failed
A manual search for a “painting of woman with scroll” would never have found it. The work stolen twenty years earlier showed the same fair-skinned woman clutching a small white lamb and a palm branch — not a scroll. Somebody had repainted her hands.
The system matched it anyway, because it was not searching words. It compares hidden visual fingerprints: shades of colour, rough silhouettes, the arrangement of a composition. It flagged the painting among millions of records despite the alteration, and once it fetched the source image it believed matched, the resemblance was immediately obvious to the human eye.
A closer investigation confirmed it: a 17th-century, one-metre-tall painting of Saint Agnese, looted from the small church of Virgo Potens in Sulmona, near Naples — more than 500km from where it surfaced.
What the system is
Since 1969 the Italian police have kept a database of looted cultural artefacts: everything from enormous oil paintings cut from church altars at night to illegally excavated Roman and Etruscan objects. It is the largest such database in the world, fed constantly by cultural police authorities and customs, and holds more than one million items. Only some entries have photographs of the original.
Historically, officers visited fairs, antique shops and auctions and compared what they saw with those records — first on paper, later by typing keywords into an online repository. It was slow and clumsy. An officer scrolling hundreds of database photographs could easily miss the detail that mattered, and the reference images were often decades old, black and white, shot from across a room on somebody’s dining-room wall, with blown-out light or extreme contrast.
“I asked myself how technology could get around this problem,” says Rapicavoli, an IT engineer who joined the stolen art team in 2013, “at least in the massive search-and-comparison phase.”
In 2020 his team conceived the Stolen Works of Art Detection System, or Swoads. “The machine doesn’t get bored,” he says.
Since its introduction in November 2023 the system has helped recover more than 300 large plundered paintings and statues, and more than 165,000 smaller coins, medals, letters, manuscripts, religious objects, stamps and cards — worth more than €203m (£170m, $226m) in total. “There has been an exponential increase in the number of hits,” Rapicavoli says.
It also catches works chopped up and rearranged, as large altarpieces often are when sold off in separate slivers. “Regardless of whether it makes sense or not, it recognises it because it’s part of a whole that was stolen,” he says. It reconciles the art without knowing anything about its meaning.
Why this one works when AI policing often does not
It is worth putting this beside a story we ran two days ago, in which researchers warned that predictive policing hardens the bias already sitting in its data, because a system trained on records of where officers were sent simply sends them there again.
Both are AI in police hands. The difference is structural, and it is worth naming because it is the test to apply to any such system.
The task is narrow. Swoads answers one question — does this object resemble a known stolen object — rather than predicting human behaviour.
The answer is checkable. A match produces a specific claim: this is the painting taken from that church on that date. Experts verify it, provenance documents exist, and a wrong match fails visibly and immediately. A predictive crime map makes no claim anyone can falsify.
The training data is not a record of police discretion. The database is a list of things reported stolen, not a list of streets officers chose to patrol. It has its own gaps — only some entries have pictures, and plenty of theft is never reported — but those gaps do not compound each time the system is used.
A human decides what happens next. The flag sent officers to look again. It did not seize anything, and it did not accuse anyone.
The advances in Italy, in technique and in policy, are opening the way for machine learning in art theft investigations elsewhere. For any institution here weighing a similar tool — a customs authority, a museum, a land records office — those four conditions are a more useful checklist than any vendor’s accuracy figure.
