How AI is revealing lost secrets of the Roman Empire (2024)

In the spring of 2023,as he drove to his SpaceX internship in Starbase, Texas, a college student named Luke Farritor found himself riveted by a podcast. The hosts were describing a competition with an audacious goal: to read a 2,000-year-old scroll without physically unrolling it. The manuscript was part of a cache of papyrus rolls buried and carbonized by the eruption of Mount Vesuvius, which wiped out the Roman cities of Pompeii and Herculaneum in A.D. 79. If opened by hand, the scrolls would crumble into pieces, obliterating whatever message they contain. Competitors who found a way to see inside using machine learning could share more than one million dollars in prize money.

“I was like, Holy cow, I have to work on this,” says Farritor, now 22. “A lot of things about it were really compelling, the biggest one being that you’re going to potentially discover a new library from the ancient world, and that’s a big deal.”

While he focused on space travel in his day job, Farritor, a computer science major, devoted his nights and weekends to the Vesuvius Challenge—something akin to time travel. What did this scroll, from a Herculaneum villa believed to have been owned by the father-in-law of Julius Caesar, have to say?

To find out, contestants would have to develop their own programs that could interpret existing 3D scans of the coiled scroll and chart subtle physical variations to detect writing on the charred material.

Farritor first studied CT scans of sections where the carbon-based ink was imperceptible against the carbonized papyrus. Another competitor in the contest, which rewarded participants with cash prizes for sharing early results, had observed “crackle” patterns resembling dried mud but which could indicate the presence of ink. So Farritor trained a machine learning model to home in on the crackle texture.

Late one Saturday night, back at the University of Nebraska, Farritor got word that a new section of the scroll had been uploaded for competitors. He was at a party, so he used his phone to log on to his desktop and put his AI model to work on the fresh image. A few hours later he checked his phone and saw the Greek letterspi, omicron,andrhoglimmering at him from across the millennia.

“Realizing, Wow, I just automatically discovered three new letters of writing from the Roman Empire, that was a pretty cool moment,” Farritor says. “I freaked out; everyone else freaked out. I sent it to the organizers. I sent it to my mom.”

(These ancient texts were once unreadable. Now technology is decoding them.)

The letters turned out to be part of the Greek word for “purple,” making Farritor the first person to glimpse a word of this scroll since it was buried under dozens of feet of volcanic debris.

The scroll was one of about a thousand recovered in the 1700s after well diggers found the villa, which lies beneath the modern town of Ercolano. Scholars say this trove was likely the Greek library of the Epicurean philosopher and poet Philodemus. But they suspect a larger Latin library may be in a section of the villa that has yet to be excavated. If the competition demonstrates that scholars can read unopened scrolls safely and easily, it could galvanize the search for more historic material.

Such a collection might contain lost masterworks of Greek and Roman history, philosophy, and literature. If no more scrolls are discovered, “this will still change our understanding of the ancient world,” says Garrett Ryan, a historian and podcaster who studies ancient Greece and Rome. But if the library is located, it will be “a profound revolution in the classics.”

(10 million enslaved Americans' names are missing from history. AI is helping identify them.)

What good is a libraryif it can’t be read? Early attempts to open the fragile scrolls damaged many of them, leaving fragments for papyrologists to piece together. Brent Seales, professor of computer science at the University of Kentucky, co-founded the Vesuvius Challenge with tech investors Nat Friedman and Daniel Gross. Seales and his lab did seminal work on ink detection and digital unwrapping that competitors are building upon. By launching the contest, they seem to have added more urgency, collective brainpower, and support for this sort of noninvasive historical investigation.

Farritor’s “purple” breakthrough was just the beginning. To make the most progress possible, Farritor and two other competitors teamed up. They went on to win the grand prize of $700,000 earlier this year for revealing 15 columns of text from the scroll, more than 2,000 characters. Their submission far exceeded the challenge’s original goal of reading four passages of 140 characters each.

All three were using AI a little differently. Like Farritor, team leader Youssef Nader, 28, an Egyptian working on his Ph.D. in AI and machine learning in Berlin, focused on improving ink detection. Rather than start with a manual inspection of crackle patterns, Nader trained his model on ink found on fragments of damaged scrolls. He applied the model to sections of a CT scan showing the target scroll’s internal structure that the contest organizers had made by placing the scroll in the beam of a particle accelerator. Nader’s process enabled crisper, more precise renderings of the characters.

(These manipulated photos are the original political deepfakes.)

Team member Julian Schilliger, 29, a computer scientist in Basel, Switzerland, worked on how to digitally “unroll” the scroll, whose layers were visible in thousands of cross sections of the 3D scan. Individual layers had to be traced through the cross sections and digitally “flattened”—its own labor-intensive procedure. Schilliger developed a method using AI to help automate the process. He created a model that identified which points in the scan were adjacent on segments of the scroll layers. “That was like a big aha moment for everyone, that you can do it automatically,” he says.

Enthralled papyrologists determined the columns the team revealed are part of a meditation on pleasure, likely written by Philodemus. One fragment reads, “in the case of food, we do not right away believe things that are scarce to be absolutely more pleasant than those which are abundant.”

As the challenge nears the end of its second year, organizers have set new goals: to read 90 percent of four scrolls, up from 5 percent of one scroll in 2023, and to scale up the techniques pioneered in the first year. Increased automation will be crucial to reading the roughly 300 intact scrolls, and perhaps thousands more. “I absolutely believe that this problem will be solved and we’ll be scanning these things and pressing a button, and 90 percent of it will just, boom, come out,” Seales says.

The progress so far has thrilled the classics world and left scholars eager to learn what more these recovered relics might have to say. “When we were first shown the text that AI had revealed, it was a moment of utter astonishment,” says Robert Fowler, chair of the Herculaneum Society in Oxford, England, which supports conservation of the villa site. “The ability to do this now and to reach across the millennia and dip into not just a single book but a library … it’s connecting with eternity.”

(New AI may pass the famed Turing Test. This is the man who created it.)