Nearly 300 years ago, archaeologists in ancient Herculaneum, once a wealthy Roman town, discovered 1,785 papyrus scrolls in a housing complex that has since become known as the Villa of the Papyri. Located near Pompeii and just seven miles from the base of Mount Vesuvius, Herculaneum was home to thousands of elite Roman citizens seeking a seaside retreat. Though the ruins were better preserved than Pompeii, with their furnishings, lavish frescoes, intact upper floors and original wooden balconies, the scrolls discovered in 1752 remained as unreadable chunks of charred ash.
This has been due to the rapid evolution of artificial intelligence technology.
Related articles
While AI has only recently entered the zeitgeist, Herculaneum's breakthrough dates back nearly two decades, when computer scientist Brent Shields first became obsessed with what to do about “damaged books that you can't even open.”
As Shields told ARTnews, it was this question — “will it remain?” — that ultimately led him to virtually map the surfaces of the ancient scrolls and make them legible.
The Villa of the Papyrus is named for its extensive library of scrolls, owned by Lucius Calpurnius Piso Caesoninus, father-in-law of Julius Caesar. The scrolls are believed to be the only known library of their kind and size in the classical world and contain important philosophical and literary texts by prominent scholars of ancient Greece and Rome. Since their discovery, they have been housed in major institutions such as the Bodleian Library at the University of Oxford, the Institut de France, and the Getty Museum.
Herculaneum Papyri in the National Library of Naples, Naples, Italy, 2019. Photo: Antonio Masiello/Getty Images
In 2005, Shields, a professor at the University of Kentucky in Lexington, heard that institutions were beginning to digitize and digitally restore important documents in their collections. An expert in machine learning, AI, and computer vision, Shields began to wonder if he could use these tools to “virtually decipher” the Herculaneum scrolls.
After four years of research, Shields completed a working prototype and convinced the Institut de France to provide the first scrolls on the condition that a system be built locally. However, that first attempt was limited by the technology available at the time, as the resolution was not “ultra-fine” and the “size of the data” exceeded their computing capabilities.
“When we came back from our first scanning session, we didn't have a computer that could actually load the whole thing into memory,” Shields said, explaining that the data was so huge that it wasn't possible to view the images of the scrolls all at once. “Now, we have all this cloud computing, and we can throw the data into systems that have lots of memory.”
Despite early advances, the technology never quite achieved the “virtual unsealing” that Shields envisioned, but by 2015, the development of cameras capable of taking tomographic images (specialized X-ray images) with such high spatial resolution led experts to conclude that the Herculaneum papyri still contained writing.
In another project, Shields and his team successfully tested new techniques using X-ray imaging and computer vision on the Ein Gedi Scrolls, discovered in an area west of the Dead Sea in Israel. Without opening the scrolls, they discovered they contained text from the book of Leviticus, a well-known document in both Judaism and Christianity.
Shields was able to confirm, by the same “virtual unwrapping” method, that the Herculaneum scrolls contained text, but the ink used was made from carbon and was not chemically different enough to be deciphered from the burnt papyrus. The densities of the ink and the papyrus proved to be too similar to be read using X-ray technology. This observation led to the decision to use AI as an intermediary to enhance the ink and make it legible, which resulted in the creation of a competition called the Vesuvius Challenge.
(It's worth noting that in addition to advances in technology, there have also been improvements in conservation methods, which have influenced the willingness of collections to allow experimental processes to be carried out on their priceless artworks and artifacts, in projects such as the Rijksmuseum's Operation Night Watch. When Shields first conducted his experiments, it was much harder to find institutions willing to have their holdings scanned.)
Interior of the House of Neptune and Amphitrite mosaic in the Villa of the Papyri, Herculaneum, Italy, 2003. Photo by Eric Vandeville/Gamma Raffo via Getty Images
In the fall of 2022, former GitHub CEO Nat Friedman heard about Shields' research and proposed an open competition to advance the research. Shields was initially hesitant, but after failing to raise research funding, he agreed. Friedman and Daniel Gross, an entrepreneur who has invested together in the AI field, put up $125,000 to launch the Vesuvio Challenge. Another $1 million was raised from other investors in Silicon Valley and social media users. In the challenge, Shields provided participants with his software and high-resolution scanned images, and they were asked to create a machine learning model that could distinguish between text and carbonized ash.
Since then, the Vesuvius Challenge has awarded prizes to participants at different stages of the competition, with each round having a specific goal. In one round that ended in July 2023, a $100,000 prize was awarded to 10 winners who competed among thousands of participants to see who could improve an AI method to detect ink on scrolls. Last year's grand prize of $850,000 was set for reading four passages of at least 140 characters each by the end of 2023.
The research, code, and methods of each round were made available to participants, allowing them to build on each other’s work.
The challenge reached a major breakthrough last October, when US physicist and entrepreneur Casey Handmer noticed a cracked, muddy texture forming Greek letters in the scans. University of Nebraska-Lincoln computer science undergraduate Luke Fariter then used the texture to develop a machine learning algorithm to identify the word porphyras (purple), winning an individual prize for the discovery. Berlin-based PhD student Yousef Nader then created a clearer image of the letters.
In the end, 18 submissions were received. After a panel of judges reviewed the code, 12 were submitted to a committee of papyrologists, who assessed readability and transcribed the text. Only one team, consisting of Falliter, Nader, and Swiss robotics student Julien Schilliger from the ETH Zurich, met the standard of reading four passages at least 85 percent readable. For their efforts, they shared a $700,000 prize.
The translated texts included previously unknown philosophical works on the senses and pleasures that discussed music, the taste of capers, and the color purple, as well as possibly a reference to the flutist Xenophantus, who also featured in texts by ancient authors Seneca and Plutarch.
Though none of the team members could read ancient Greek, Nader told ARTnews they became intimately familiar with the writing hidden in the scrolls. “The process involved tracing the letters to create a black and white image of the ink – similar to tracing,” Nader said. “We learned how this particular scribe wrote the letters, how they drew the letters from a specific point. With the high-resolution scans, we can see where the ink got to where they started to draw the letters, and how they drew them.”
For Domenico Camardo, archaeologist at the Herculaneum Conservation Project, the Vesuvius Challenge was a real breakthrough.
“Knowing the enormous effort that papyrologists working at the Herculaneum Papyrus Workshop at the Naples National Library have put into deciphering the Herculaneum Papyri over the decades, I was amazed at how the AI managed to recognise letters, then words, and even reconstruct entire texts without risking destroying the charred scrolls by unfolding them,” Camardo told ARTnews in an email.
Text from PHerc.Paris. 4 (French Institute). Unpublished for 2,000 years. Courtesy of Vesuvius Challenge
A new iteration of the Vesuvius Challenge has been announced for 2024, with the goal of deciphering at least 90 percent of the four scanned scrolls by the end of the year. Shields said the ultimate goal is to increase the number of scrolls read while continually improving the speed and accuracy of the technology.
“Every little thing we learn has the potential to be game-changing,” he said. “You never really know when a really big discovery is going to happen. The interesting thing about Herculaneum is that each of those books could become anything.”
While these scrolls may contain influential works of literature and history previously unknown, it is important to note that they are only part of a larger library yet to be discovered. Shields hopes that these continuing developments will encourage further research at Herculaneum, since there are still parts that have yet to be excavated and the main library has yet to be identified, meaning there could be thousands of scrolls buried beneath the ash.
The new techniques discovered through the Vesuvius Challenge can also be applied to the deciphering of other texts, and their implications have not gone unnoticed.
“I've seen promising results from the Egyptian scrolls that I got for my project here in Berlin,” Nader said. Through the model of Herculaneum, “they've learned something about papyrus and ink, and that can be useful for the Egyptian scrolls,” he explained.
While many have expressed concern about the impact of artificial intelligence on human creativity and unemployment, Nader added that people in the archaeology field are beginning to see AI as “fundamentally a tool that can be used for a variety of purposes.”
In the field of archaeology, AI has already helped locate and identify many previously inaccessible finds. One example is the use of LiDAR laser remote sensing technology to discover hidden structures and pyramids built by the Mayans in the dense jungles of Mexico. Another example is the use of ground-penetrating techniques to identify a previously unknown L-shaped structure in the ancient Egyptian necropolis at Giza. AI has also proven to be a useful tool in this field in many other ways. For example, the Pompeii Archaeological Park has deployed AI-enabled robotic dogs to protect the ruins from theft.
“AI is successful because of human-shaped data, so it's interesting to see AI playing a role at the intersection of the humanities and these new technologies,” Shields says. “It's no coincidence that to build large-scale language models, we need all of the human writing, because human writing, in some sense, captures what it means to be human. That interplay is the next frontier.”
