This New A.I. Chatbot Hunts Nazi

Art & Tech This New A.I. Chatbot Hunts Nazi-Looted Art The nascent A.I. Provenance Assistant is rendering data far easier to read. The nascent A.I. Provenance Assistant is rendering data far easier to read. Vittoria Benzine ShareShare This Article The New York-based Center for Art Law estimates that between 1933 and 1945, Nazis either stole or forced the sale of 650,000 artworks. Researchers at California-based Santa Clara University have joined the quest to repatriate the 100,000 looted relics still at large by creating the A.I. Provenance Assistant, a new chatbot trained to comb convoluted records for leads. SCU management professor Michael Santoro began spearheading the project after learning of U.S Court of International Trade Judge Timothy Reif’s efforts to repatriate artworks stolen from his grandfather’s cousin, Viennese cabaret performer Fritz Grünbaum. In an article on Santoro’s project, SCU noted that previous museum-led attempts to amass data on Nazi-looted artworks resulted mostly in siloed resources. As such, the newfound A.I. Provenance Assistant—created in collaboration with information systems and analytics professors Haibing Lu and Michele Samorani—stands to save sleuths time by perusing these fragmented, error-ridden databases, which often span multiple languages, at previously impossible speeds. The crew began constructing their tool by scraping the ERR Project’s database of artworks that passed through Paris’s Jeu de Paume Museum, where Nazis once processed their loot—including a rare drawing by Rococo pioneer Jean Antoine Watteau that recently went to auction, and Baroque artist Nicolas de Largillière’s estimate-smashing Portrait d’une femme, à mi-corps (ca. 1700), which France’s famed Monuments Men recovered in May 1945. Next, Lu led the creation of an A.I. agent trained to apply simply-written queries to that database, “improving on the search by seeking out related terms, different spellings, and translated versions of names,” according to SCU, which described the Assistant as a “data docent” programmed to then explain its findings in plain English. Rather than building their own Large Language Model, the developers tailored existing LLMs to their own purposes, Lu told me via email. “Our goal wasn’t to replace provenance researchers,” he noted. “It was to make decades of complex archival records much easier to explore through natural conversation.” The A.I. Provenance Assistant is already live. Nevertheless, Santoro’s team plans to keep improving the tool, namely by expanding upon its databases. The crew is also gearing up to publish a paper once the project reaches scalability—and establish a related foundation capable of hiring lawyers and raising money for the presently unfunded endeavor. To that end, Santoro has recruited tech and corporate strategist Wendy Goldberg to help articulate the project’s potential. “When I was at AOL, it was all about making the internet as easy to use as the telephone and the television,” Goldberg told SCU. “This is making using A.I. for restitution and tracking art as easy to use as any other technology application. You don’t have to be a genius to find the provenance and track the trail of stolen art.” The developers are also seeking input from expert provenance researchers, in order to learn how the AI Provenance Assistant can better suit their needs. “We want to come up with tools that help researchers identify high-risk works, to give them something to investigate,” Lu told the university. “In some sense, we cannot even envision who is going to use it.”