AI just deciphered part of an ‘unreadable’ ancient scroll. Here’s what it says.
The Herculaneum scrolls were so badly damaged in the A.D 79 eruption of Vesuvius that scholars feared the ancient library was lost forever. That just changed—with help from technology and a $1 million prize.
With the help of artificial intelligence, researchers are now able to read 2,000-year-old passages of text from a scroll charred in the A.D. 79 eruption of Vesuvius that was long considered unreadable. The breakthrough was announced Monday by the organizers of the Vesuvius Challenge, a global competition offering more than $1 million in prize money for significant benchmarks in applying machine learning techniques to coaxing long-lost texts from the famed Herculaneum scrolls, an ancient library assumed to be lost forever.
Roughly 1,800 papyrus scrolls—believed to contain literary and philosophical works from the first and second centuries B.C.—were discovered by workers digging up the ancient town of Herculaneum near Pompeii in 1752. The scrolls had been reduced to brittle, charred lumps by the heat and gasses of the eruption, and the ones that workers didn’t throw away more than 250 years ago have largely languished since then in storerooms, written off as indecipherable curiosities.
Technological developments over the past two decades have helped researchers get closer to being able to “read” the fragile scrolls. But only the very recent acceleration of artificial intelligence and computing have finally made it possible to begin unlocking their secrets—all without opening them.
This week, the Silicon Valley investor-backed Challenge awarded its latest installment, $700,000, to a team of three competitors—an American college student, an Egyptian graduate student in Germany, and a Swiss robotics engineer—who worked together to reveal 15 columns of text totalling more than 2,000 characters from an intact scroll.
Epicurean delights
Papyrologists have produced a preliminary transcript of the newly revealed text, which represents about 5 percent of the scroll’s content. They’re still working to analyze the text, but shared a few snippets that amount to a “2,000-year-old blog post about how to enjoy life,” according to the contest organizers.
Based on what they’ve gleaned so far, papyrologists suspect the unnamed scribe was Philodemus, a follower of the Greek philosopher Epicurus, who valued pleasure above all else. One passage, for instance, contemplates how abundance or scarcity might affect sources of pleasure like music and food. “We do not right away believe things that are scarce to be absolutely more pleasant than those which are abundant,” the author writes.
Historians suspect Philodemus was the philosopher-in-residence at the Herculaneum villa where the charred scrolls were found—a villa likely owned by Julius Caesar’s father-in-law. The deciphered text mentions Xenophantos, for instance, who may be the same musician Philodemus notes in other writings. The author also criticized his adversaries—likely the Stoics—for having “nothing to say about pleasure, either in general or in particular, when it is a question of definition.”
‘The scrolls are readable’
The newly revealed passages build upon “πορφύραc,” a single word from the scroll, which two of the three winning competitors revealed independently in October. The colorful ancient Greek word refers to purple dye or purple-colored clothes—a color closely associated with royalty and power.
“We knew if we could read just one [scroll], then all the other ones would be available with the same method or some augmented method,” says Brent Seales, a computer scientist at the University of Kentucky who’s been trying to decode the Herculaneum scrolls for the last 20 years and leads the university’s Digital Restoration Initiative, said in October. “We are now proving not just to ourselves but to the entire global community that the scrolls are readable.”
As they progress in deciphering the Herculaneum scrolls, researchers suspect they’ll find many more Epicurean reflections. One scholar likened the villa’s Epicurean-obsessed owner to “the sort of guy who collects every album of one band and every cover of that band, but nothing else.”
But, beyond that, all scholars can do is make educated guesses—and draft their wishlists. Perhaps the scrolls will contain other kinds of Greek philosophy, such as works by Aristotle or Stoic texts. Or, maybe, some Latin literature or the long-lost writings of authors like Sappho, Homer, and Sophocles.The scrolls might also offer fresh insights about early Christianity, some researchers suggest. But only time—and more virtual unwrapping and transcribing—will tell.
Reading the Herculaneum scrolls
Since the mid-1700s, people have made various attempts at reading some of the less damaged scrolls from Herculaneum. One method involved cutting the scrolls in half and scraping away layers one at a time to see the text inside; another involved slowly unwinding the scrolls with a specially built machine. Though these 18th and 19th century efforts did allow conservators to copy down some of the words inside, they often damaged—or, worse, totally destroyed—many of the scrolls in the process.
Between 500 and 600 carbonized scrolls from Herculaneum—kept in museum, university and national collections in England, France and Italy—remain unopened, though the exact figure is hard to estimate because many are fragmented. The remaining scrolls are extremely brittle, which means physically unrolling them is not a viable option. “If you drop one, it would shatter like glass,” Seales explains.
Technological advancements since the early 2000s have helped researchers overcome this hurdle, including using CT scans to make 3D images of ancient scrolls. From there, Seales’ Digital Restoration Initiative team developed software that could “virtually unwrap” the 3D images to produce flattened segments. This method enabled them to read previously hidden text from the Ein Gedi scroll, a charred and fragmented scroll from the Middle East dated to the third or fourth century A.D.
When researchers tried to use this method to read the scrolls carbonized by Vesuvius, however, they ran into another roadblock. The ink used on the Ein Gedi scroll contained metal, which meant the letters were visible on the CT scan. The Herculaneum scrolls, by contrast, were written with carbon-based ink, which, to the human eye, makes the symbols indistinguishable from the carbonized papyrus on the CT scans.
Undeterred, researchers wondered if higher-resolution scans of the scrolls produced by a particle accelerator could provide an even more detailed view of the carbonized papyrus. Sure enough, at very high resolutions, the scans revealed visible areas where the ink slightly altered the shape and texture of the papyrus fibers. “The carbon-based ink sort of fills in the holes that are the grid of the papyrus—it coats them and makes them a little thicker,” says Seales.
Seales and his Digital Restoration Initiative colleagues then developed and trained a machine learning model to detect these subtle differences in the carbonized papyrus surfaces. But to take the project any further, they needed human beings to help. That’s where the Vesuvius Challenge comes in. Hoping to harness the collective power of citizen scientists around the world, Seales teamed up with Silicon Valley investors and put his team’s data, code, and methods online for anyone to access. The challenge’s pitch? After 275 years, the puzzle of the Herculaneum scrolls has been reduced to a software problem—one that anyone, anywhere with access to a computer could, in theory, contribute to solving.
Citizen science
In March 2023, the Vesuvius Challenge team released thousands of 3D images of two rolled-up scrolls from Herculaneum, as well as a machine-learning algorithm trained to detect the invisible letters and symbols written on the layers of carbonized papyrus. They also offered $1 million in prize money to incentivize participants to build upon the AI technology and, ultimately, speed up the deciphering.
Three competitors extracted the new columns of text working collaboratively: Luke Farritor, an undergraduate at the University of Nebraska-Lincoln; Youssef Nader, a doctoral student at Freie Universität Berlin; and Julian Schilliger, who recently earned his master’s degree in robotics from ETH Zürich. They split the 2023 grand prize of $700,000 for being the first competitors to reveal at least four separate passages of 140 characters each, with no more than 15 percent of the characters missing or illegible, by the end of 2023.
All three had won earlier prizes in the contest. Farritor and Nader shared the $50,000 “first letters” prize for separately revealing πορφύραc (Farritor got $40,000 for being first, while Nader got the remaining $10,000). Schilliger, meanwhile, won three smaller progress prizes totaling $14,500.
Though the 2023 grand prize has been awarded, the contest is far from over. Organizers have announced new incentives for 2024, including $100,000 for the first person or team who can reveal 90 percent of the text in each of four scrolls provided by the Challenge. This, organizers believe, will lay the groundwork for eventually being able to read all of the ancient texts.
Citizen scientists can find everything they need online, from the history of the scrolls themselves to downloadable data, algorithms, and tutorials. And while the contest is open to anyone, it’s technical work that’s so far mostly attracted computer scientists who are already well-versed in machine learning. Competitors are helping advance the project forward by virtually unwrapping additional sections of scrolls via software and methods developed by Seales; they’re also working to improve the machine learning model by providing it with additional training examples from the newly unwrapped digital segments of papyrus.
Curiosity and prestige
The competitors—an estimated 1,500 to 2,000 of them in total, according to Seales—have done their part. In less than a year, they’ve made huge strides toward solving a puzzle that has stymied people for centuries. “We’ve seen 10 or 20 person-years of work from these competitors,” says Seales.
So, what’s motivating the contestants to volunteer hours and hours of their time toward the project? The prize money is a big factor but, on top of that, some competitors are simply intrigued by the scrolls themselves. “When things were a bit frustrating and things were not working, I felt like I was unable to give up because I was just too curious—I really need to know what’s going on here,” says Nader.
There’s also the allure of working on a project backed by Silicon Valley entrepreneurs and investors. Former GitHub CEO Nat Friedman launched the contest, along with venture capitalist Daniel Gross; other startup founders and investors also chipped in prize money. “There’s kind of this Silicon Valley prestige,” says Farritor, who spent last summer interning at SpaceX.
From here, the machine learning model should continue to improve even more and reveal additional letters until, ideally, researchers will be able to decipher all of the Herculaneum scrolls. These efforts could pave the way for future excavation work at Herculaneum, where experts believe even more scrolls are still buried—and more lost texts from ancient writers await.
“Some people might think, ‘What are you going to all that trouble for?’ but I don’t believe that,” says Seales. “This is an amazing period in human history. We’re talking about more works from that period. Yeah, I want more, I want it all.”
