How to Catch a Criminal in the 21st Century and Why AI Might be Able to Help

In real life, art crimes are less glamorous and leave much greater wreckage in its wake, including tarnished reputations, heartbroken owners or decimated cultural heritage sites. The paths to recovering such pieces are numerous and multifaceted and at times protracted, particularly if the piece itself is ancient or has never before been cataloged, leading to a much more complicated and potentially mistaken provenance. ^{As a way to combat these ongoing issues, databases became a popular way to try and retain information on stolen works, including ownership history, descriptions, frame information, and images.}

Repatriation and restitution within international law are governed by a series of treaties with the most recent being created during the UNIDROIT Convention of 1995. ^{This treaty specifically calls for a uniform law to be implemented for the return of stolen cultural heritage, but it remains unsigned by important countries such as the United States. Therefore, the 1970 UNESCO Convention remains the agreement primarily at play. Unlike the UNIDROIT document the UNESCO Convention doesn’t create a singular system that must be enforced, but instead looks to create a framework to help parties create legislation protecting cultural heritage. To read more about these conventions check out our Art Law Lunch Talk from 2020 “UNESCO Convention Turns 50”.}

In more recent history, the Nicosia Convention created a new treaty which imposed criminal penalties on certain actions like illegal import, or the falsification of provenance documents. ^{This treaty entered into force in 2022 once the fifth signatory voted for ratification. Only six states have chosen to ratify the document. To read more about this treaty and the criminal charges it imposes read our article, “Revisiting the 2017 Nicosia Convention.”}

Law enforcement around the world including Interpol, the Carabinieri and the FBI maintain their own versions with more tightly held controls, but still allow for a public facing element that includes a free mobile App. Check out our article by Eleanor Gartstein titled “Stolen Art Databases, Bridging Gaps, and Balancing the Need for Private Policing” to read more about the current limitations of decentralized databases and “Scrolling Through Antiquities: INTERPOL ID-Art App” by Claire Darrow to find out more about Interpol’s App in particular.

Private detectives and non-profits, including the Center itself, have taken it upon themselves to create and manage some of this information, storing it within cyberspace. The leading private databases, including the Art Loss Register (“ALR”), which contains over 700,000 entities, allow individuals to conduct a search for a fee. ^{An outgrowth of ALR, The Cultural Heritage At Risk Database (“CHARD”), manages a database of works, specifically movable objects, that are located in conflict zones to help private and public organizations with identification of stolen works and stop illicit sales.}

The International Council of Museums keeps what they have named the Red Lists Database. ^{A system of booklets that they distribute to law enforcement in the hopes of maximizing the cooperation and support to stop illicit trafficking of goods. Currently, the Council keeps lists covering areas most at risk for including Ukraine, Iraq, Cambodia and more.}

These decentralized repositories are only as good as the search capabilities of the individual using the system. And although there are programs looking to promote research skills (check out BellingCat’s Open Source Challenges), the average customs official is not versed in deep web and social media listening. If an official is lucky and the work does exist within a database, and even luckier it exists within the database they were able to search within, they might not have the experience or knowledge to identify the signs of a faked provenance or looting marks. In the search for new solutions Artificial Intelligence has come to the forefront as a potential solution.

As opposed to the negative implications regarding art and AI in the copyright world, and the ongoing litigation regarding the use of copyrightable works in AI training and the generation of copyrightable work, AI is emerging as the next big thing in the search for tools to help find stolen cultural heritage (for more on AI and copyright check out this article and our prior event). Similar to the numerous databases extant to track stolen artwork there are many ongoing projects both implemented and still being created across the world. Below is a selection of the various tools in development.

The commander of the Carabinieri has stated that in 2023 105,474 pieces worth more than $287 million were found by SWOADS. ^{The Carabinieri also have stated that they wish to make SWOADS available to any countries who would like to use the platform, with the hope that those countries will also help contribute to the database. For their impressive work with SWOADS (not to be confused with swords), the Carabinieri received the Innovative Police Force Award in 2024 at the World Police Summit. SWOADS has the potential to be the global benchmark in how to effectively incorporate AI in the fight against the illicit art market.}

German App KIKU

The plan was for a customs official to take and upload images of imported objects into KIKU. ^{The app would then use image-recognition software to provide initial information like the origin of the work and its potential date of creation, after which it would search stolen art databases to find a potential match, alerting the official if one did appear. If no match were found, the information provided by the app could still help officials verify if the provenance documentation accompanying the object appeared false (forged, incorrect, etc) . Initial publications surrounding the development of the app mention various limitations including that the AI uses the Prussian Cultural Heritage Foundation images and data which only contains around 2,500 items, which is not enough for a robust database; the lack of information on a newly looted object is a second limitation. In coordination with the app, Germany funded NEXUD, a group of antiquities experts at the Cultural Foundation of the German Federal States, to assemble a group of experts to work with KIKU’s findings. As of July 2025, no updates could be found on the current status of the app, either in regards to its implementation or lack thereof.}

Boston University’s Khmer Statuary Project

In the spring of 2025, archeologists from Boston University introduced a new database called the Khmer Statuary Project (KSP). ^{KSP uses machine learning to identify stolen Cambodian statues. The BU team trained KSP using 690 images from various databases and collections. The database identifies a user imputed image providing 12 potential matches, thereby reducing the workload for researchers trying to determine if a Cambodian in question statute is looted.}

Similar to KIKU, the database is severely limited in its capabilities due to the size of the training pool. The researchers do believe the capabilities could be expanded to include artwork from other regions at risk for looting. ^{They chose to focus on Cambodia first due to the efforts by the Cambodian government to repatriate looted pieces.}

The history of looting in Cambodia traces its roots all the way back to French colonization, but one of the most notorious smugglers was Douglas Latchford who was indicted in 2019. ^{After previously being seen as a collector and scholar of Cambodian art, including writing several books, Latchford fell from grace when a Sotheby’s statue he was connected to was re-examined and determined to be looted. This investigation turned up a cache of documents now called the “Pandora Papers” which exposed a collection of multiple artifacts supposedly tracing back to an offshore trust allowing for the obfuscation of the items provenance. This ended with at least 27 artifacts from prominent museums to be linked back to Latchford’s looting. In an ironic twist of fate, the books Latchford helped to author could be used to train the AI that will find the pieces he stole.}

In the world of art forgery detection and authentications, a team from the University of Oregon published new research regarding their use of machine learning to identify fake Jackson Pollocks with a 98.9% accuracy rate. ^{Alternative to the more commonly developed methodologies of facial recognition, color, or other artist specific techniques, the UO team trained the AI based on 588 works arranged into tiles based on fractal patterns, therefore expanding the original data set to 97,275 Pollock tiles and over 150k non-Pollock tiles. The trained AI then appears to look at the scale invariance of poured work to determine if it matches that of a Pollock, or not.}

The researchers have stated that at this juncture the specialized version of authentication through fractals will not work for artwork that is not poured, although they do believe that it might help with other gestural art. ^{The researchers themselves envision the use of multiple diverse machine models used together for authentication. One member of the team Professor Robert Taylor has stated in regards to the computer ““[o]ur computer can spot a fake far more accurately than a human,” he said. “Is that a form of artistic appreciation? In a way, artificial intelligence does appreciate a Jackson Pollock painting.”}

The Significance project was funded by the European Union to research AI’s capability to monitor and detect stolen cultural heritage goods, specifically through a combination of image and pattern recognition, when fed information received from the internet.

The project employed a Convolutional Neural Network (“CNN”) to look at an image and determine if a cultural heritage good is present, what the goods classification is (painting, coin, ect.), and pull information regarding the good including location and period. ^{The CNN does this by being fed multiple images and subtracting specific data features then using those features to compare and distinguish new images based on particular objects. The study found that there was a high level of accuracy achieved which could allow for a streamlined process and more effective tools to track illicit trade in cultural heritage goods.}

If this is of further interest don’t miss our “Some Like It Digital Webinar”coming this fall about responsible uses of AI in authentication.

Conclusion

In the world of illicit cultural heritage goods there is an overwhelming amount of missing goods. AI has been found to be a useful tool able to reduce some of the work load through the use of image and pattern recognition. In the use case of authentication there become questions regarding the efficacy of the images the AI has been trained on due to the potential prevalence of inauthentic works being considered authentic, therefore skewing results. This is less of a concern in the tracking of illicit goods, but both markets contain limitations due to the number of images available for an AI to train on. These complications come both from the outstanding legal question of training an AI on copyrighted works, transparency in datasets and techniques used to train models, and the lack of imaging on items that are stolen or smuggled.

For all sources in general, AI, although a powerful tool with an impressive ability to churn through datasets in minutes that would take a normal team months, it is not an infallible system. For every AI developed there are certain parameters created and margin of errors that a researcher deems within the normal allowance. Suggesting that Germany’s choice to devote a team to helping analyse the information, might be a necessary component to perfect the implementation.

In regards to the continued training of an AI solution, whether centralized or not, organizations should look into potential licensing deals with privatized databases such as the ALR or CHARD. The implications of the use of a private database comes with further questions regarding cooperation and monetary payment from a government agency. There are also issues with what information should or should not be kept private both for the victims and for the private institutions data protections.

Although the multitude of ongoing research is impressive and promising, if one centralized accessible network fails to emerge, law enforcement will be hindered due to jurisdictional boundaries and inaccessible information, or expected to conduct redundant searches in multiple repositories. Similar to the division between the various databases that already exist, without international cooperation, the evolution of AI tools will remain a piecemealed conglomeration limiting the success of what is possible.

This is a definite worry as political tensions, nationalism, and xenophobia is on the rise within the US, France, the UK, Germany, and beyond. ^{Hopefully cooler heads will prevail, with a built in option like SWOADS already in existence a best path forward would be for other groups such as Interpol and the FBI to work in collaboration with the Carabinieri to build upon their current system while combining all of the possible government managed databases.}

Shelby Jorgensen is a rising 2L at the University of Wisconsin Law School, working as a Summer 2025 Legal Intern for the Center for Art Law. A 22’ graduate from the University of Notre Dame with a dual degree in marketing and studio art, Shelby hopes to combine her love for art with her interest in the law to work as an intellectual property attorney. She can be contacted for questions or comments at sjorgensen4@wisc.edu.

Select Sources:

Facebook
Twitter
Pinterest
LinkedIn
Tumblr
Email