Carabinieri Command for the Protection of Cultural Heritage
Art theft is a global problem and police forces are struggling to keep on top of it. Now they are turning to AI software that snoop out the most cleverly hidden artworks.
During their routine visit to the annual antiquity fair in Parma, Italy, in October 2024, police officers busily snapped photos of the artworks as they walked between the stands. It wasn’t art appreciation – it was detective work.
The images were uploaded to an internal database as the officers smiled at the vendors and wandered amid fist-sized bronze statuettes, patinated goblets, yellowing globes, and an endless assortment of oil paintings.
Just a couple of hours later, though, the police system flagged one of the snapshots with an alert signal: a painting of a young, fair-skinned woman wearing aquamarine and red, holding up a half-unravelled parchment scroll.
The next day, officers rushed to the scene and seized the painting for their art experts to take a second look. The scroll in the painting was now, indeed, giving the investigators reason to be suspicious, says Lt Col Salvatore Rapicavoli, of the police unit for the Protection of Cultural Heritage of Italy. “There was something off.”
While Rapicavoli’s team has specialised in tracking down stolen artworks nationwide for decades, he says that the novel, tailor-made artificial intelligence system they were testing at the fair has massively sharpened their craft. It has since allowed the team to detect stolen art more quickly and helped them to spot details beyond what the human eye can normally see.
Carabinieri Command for the Protection of Cultural Heritage
“There has been an exponential increase in the number of hits,” says Rapicavoli. Since its introduction in November 2023, the AI system has helped his team 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 a total of €203m (£170m, $226m). The advances in Italy – in technique and policy – have been paving the way for machine learning in art theftinvestigationsworldwide, promising to revolutionise the global crackdown on art heisters.
And while a human would likely have glossed over that woman holding a parchment scroll at first glance – nothing but a couple of darker shadows looking off balance – the team now had an active investigation on their hands.
A huge haul
Since 1969, the Italian police have kept a database of looted cultural artefacts across the nation – everything from gigantic oil paintings ripped from church altars in the night to ancient Roman and Etruscan artefacts excavated illegally. The database – the largest in the world, constantly fed by reports from cultural police authorities and customs – contains more than one million items. Though only some of the entries have pictures of the original artefacts.
Historically, to find stolen artworks, Italian police routinely visited fairs, antique shops, and upscale art auctions and compared what they saw there with millions of records back in their offices, first in paper form, then digitised in an online database. One by one, they would look for matches in the online repository by searching keywords, querying with details like “painting of woman with scroll”.
This was time-consuming and clumsy, though, says Rapicavoli. Even if a sale looked suspicious, an officer scrolling through hundreds of database photos for a match could often miss key details. And the quality of the original artefact photos left a lot to be desired. The images were often taken decades earlier, in black and white, far away, on someone’s dining room wall and littered with light blowouts, obscured areas, or extreme contrast.
The AI recognises it because it’s part of a whole that was stolen – Salvatore Rapicavoli
“I asked myself how technology could get around this problem,” says Rapicavoli, an IT engineer who joined the Italian police’s stolen art team in 2013. “At least in the massive search-and-comparison phase.”
In 2020, his team conceived the AI algorithm called the Stolen Works of Art Detection System (Swoads). “The machine doesn’t get bored,” says Rapicavoli. Crucially, it can match hidden visual fingerprints in the works – pairing similarities between shades of colour and rough silhouettes.
It can also spot segments of paintings that have been chopped up and rearranged – as is often the case with large altarpieces sold in separate slivers – or works that have been tweaked after being stolen. “Regardless of whether it makes sense or not, it recognises it because it’s part of a whole that was stolen,” says Rapicavoli. It can reconcile the art without truly knowing its meaning or content.
Themystery of Saint Agnese
A manual search for a painting of a “woman with a scroll” in the stolen artworks database wouldn’t have solved the case on Rapicavoli’s hands. The keywords wouldn’t have surfaced the painting stolen 20 years earlier, in which that same fair-skinned woman clutches a small white lamb and a palm branch instead of a scroll.
Because the AI had spotted the similarities in colour, pattern and set-up – irrespective of the missing scroll – it flagged it among millions of stolen artefacts despite the manner of its disguise. Once it fetched then eye were immediately apparent
In fact, a closer investigation confirmed that a 17th-Century, 1m-tall (3.3ft) painting of Saint Agnese had been looted from the small church of Virgo Potens in Sulmona, near Naples, over 500km (310 miles) from the Parma art fair, on the night of 11 August 2004. It had apparently then been painted over to appear less recognisable – the lamb now the scroll, and the halo and the palm branch erased to turn the Saint into a young pagan woman instead.
Carabinieri Command for the Protection of Cultural Heritage
As the art trade moved from in-person auctions to online marketplaces, Rapicavoli’s team was finding itself constantly playing catch-up with a flurry of new uploads across e-commerce platforms, social media, and even the dark web. Now Swoads can continuously crawl catalogues and new uploads day and night, constantly scraping the internet and comparing millions of images against a stolen-art database.
In 2026, for instance, the same software allowed his team in the Italian city of Turin to recover a painting by Filippo Palizzi, Riposo Della Pastorella (The Rest of the Shephardess), stolen from a modern art gallery in June 1972, in a robbery so severe the establishment remained shut down for over 30 years.
Although the painting had been missing for decades, lost in the backrooms of art markets and trade shows, one of its routine trawls of the internetspotted it for saleon an auction house website.
An aid, not a replacement
Still, AI isn’t a silver bullet against art theft, says Carina Popovici, the co-founder of Art Recognition, a company using AI for art authentication. These systems answer a specific question: does this image match something reported stolen? They don’t catch the thief, and they can’t flag something that was never recorded, she notes. “The recovery still depends on investigators and legal cooperation,” says Popovici. “I would call AI a force multiplier.”
Timothy Carpenter, the former head of the FBI’s Art Crime Team, and founder of Argus Cultural Property Consultants, a risk management and consulting firm, agrees. The software can’t replace human analytical thought.
You have to have a large team dedicated to nothing but chasing leads generated by the AI – Timothy Carpenter
For one, there are still many technical issues to iron out. While the technology has become much better at recognising stolen art even from low-quality photographs, it still struggles with poor training data. Its recognition of 3D objects, too, is severely impaired – a vase photographed from the front in the theft report will likely not match a listing photographed from the back. Rapicavoli’s team is already working to improve this.
And there are systemic issues: “You have to have a large team dedicated to nothing but chasing leads generated by the AI,” says Carpenter – a luxury most law enforcement branches cannot afford. “The Italians have somewhere in the neighbourhood of 300 personnel assigned to their art crime unit. At its largest, the FBI had 24.”
AI also can’t necessarily distinguish a copy from the supposed original: many copies of an artwork can exist in the world, whether authorised or not. Because these databases don’t require proof of authenticity before adding a stolen work, says Carpenter, the stolen object might not even be the only existing version of a painting. For instance, Interpol’s database currently lists that Murillo’s The Melon Eaters was stolen, says Carpenter: “But the original painting still hangs in the Alte Pinakothek [in Germany] and is not reported stolen.” Interpol confirmed that’s because the database instead records the theft of other versions of this painting: a copy attributed to the artist, reported stolen in Rome in 2009, and a copy by a disciple, falsely signed by Murillo, reported stolen in Bucharest in 1998.
Interpol said in a statement to the BBC: “In the case of the Murillo painting mentioned, we acknowledge that an authenticated original is currently conserved at the Alte Pinakothek in Munich. However, it is common for multiple versions of a masterpiece to exist, including copies created during the artist’s lifetime, works by disciples, or pieces attributed to the master, etc.”
Carabinieri Command for the Protection of Cultural Heritage
Talk about training AI to undermine forgery, in fact, abounds. The algorithms could spot duplicates or suspicious prices, or work to examine imperceptible details such as brushstroke geometry, colour relationships, and texture. But this is a slower and more intricate process to set up with the tools available now, says Steven Frank, an intellectual property attorney who researches how to use AI to authenticate art.
Not to forget that the same AI tools helping investigators can also help the criminals, allowing them to manipulate an image, change a frame, digitally repair damaged areas on a painting, and, of course, fabricate false provenance documents or create realistic synthetic objects that don’t actually exist, says Carpenter. “So, AI is making things easier, while at the same time making them harder,” he says.
International trafficking
All of this is on Rapicavoli’s mind. His team will update the software with watchlist data on artefact categories forecast to hit the market, such as rare Ukrainian manuscripts or icons likely to be looted during the conflict, so the system can stay on high alert. They are also working toward an internationally interconnected identification network by 2027 – given that the Italian’s work hardly exists in a silo.
Turkey has its own AI system, which it has already, for instance, used to recover various stolen, famous 15th-Century tiles from the UK. French researchers have also created their version of the app, Arte-Fact, and German cultural institutions have launched one called Kiku. In the US, Boston University researchers are using machine learning algorithms to identify statues that may have been stolen from Cambodia. Researchers at Santa Clara University in California, meanwhile, have created an AI Provenance Assistant to help scrape records of art looted by the Nazis.
Interpol, ultimately, has software called ID-Art, connecting databases on a global scale. Thanks to its image-recognition abilities, the Spanish Guardia Civil recently recovered two stolen Renaissance tables more than 40 years after their theft by matching photos against blurry black-and-white database images from 1979. They also recovered three Spanish Roman Empire gold coins worth up to €200,000 (£167,600/$224,000) .
Not‘mostwanted‘
No matter how advanced, AI will always likely be less useful in recovering more recently stolen works, or more high-profile thefts, says Frank. “The thieves know better than to try to sell them through traditional channels,” he notes. Many of these thefts are directly commissioned by a collector, and will probably enter a vault or freeport storage facility.
• The secret codewords that scammers hate
• AI has ruined cute animal videos
• Your car is spying on you, and it’s only just the beginning
Most cases chalked up by the new AI systems are, in fact, lesser-known pieces, and the thieves are often willing to devalue the work – chopping it up, modifying it – rather than risking getting caught. “I don’t expect the system is going to find me the Antonello Da Messina works that were stolen this summer, let’s be very clear,” says Rapicavoli. “I don’t expect it to find me Caravaggio’s Nativity, one of our ‘Most Wanted’.”
For instance, there is no evidence to turn the Sant’Agnese painting theft into a real criminal case. The theft happened over two decades ago, so Italian law can’t prosecute it. And the antiquities reseller from Naples had little information on its provenance, likely inheriting it through his family business.
But just because this isn’t a splashy, multimillion-dollar art-heist bust doesn’t mean it lacks sentimental and territorial value to the community, says Rapicavoli. On 20 May 2025, his team got to return the painting to the small Virgo Potens church in Naples, where it will be restored and hung back in its place. The locals, he recalls, were elated.
For more technology news and insights, sign up to our, whileThe Essential Listdelivers a handpicked selection of features and insights to your inbox twice a week.
For more science, technology, environment and health stories from the BBC, follow us onFacebookand<a href="https://www.instagram.com/bbcfuture_official/” rel=”nofollow noopener” target=”_blank”>Instagram.
Intelligence Revolution
Artificial intelligence
Science
Computer
Crime
Features
