A calm academic routine abruptly veered toward something unfamiliar on a fall afternoon in October 2023. An email with a blurry picture of what appeared to be a scrap of burned paper was sent to Federica Nicolardi, a papyrologist who studies ancient texts. It was not paper at all. When Mount Vesuvius buried the Roman town of Herculaneum beneath volcanic debris almost two millennia ago, the papyrus scroll was burned.
Since the year 79, the scroll had lain silent. Scholars had attempted to unlock these scrolls for centuries, typically with disastrous outcomes. The papyrus had a tendency to crumble like charcoal after being baked and carbonized by volcanic heat. Sentences were replaced by fragments in libraries. Historians frequently felt uneasy as they stood in front of those blackened cylinders, knowing that knowledge was there, visible but out of reach. Unexpectedly, artificial intelligence has now started to read them.
| Field | Information |
|---|---|
| Historical Discovery | Herculaneum Scrolls |
| Estimated Age | ~2,000 years |
| Original Event | Burial during the eruption of Mount Vesuvius (AD 79) |
| Key Research Lead | Brent Seales |
| Institutions Involved | University of Kentucky, Diamond Light Source |
| Technology Used | CT Scanning, Machine Learning, 3D Segmentation |
| AI Competition | Vesuvius Challenge |
| Prize Awarded | $700,000 |
| Key Breakthrough | Reading ancient Greek text without opening scrolls |
| Reference | https://scrollprize.org |
It’s similar to seeing archaeology done inside a computer when you watch the process take place. There is never a physical opening of the scrolls. Rather, they are scanned at facilities like the Diamond Light Source near Oxford using incredibly detailed CT imaging. A dense digital model of the scroll is created by the scans; it is a three-dimensional landscape composed of tiny units known as voxels, which are essentially the 3D cousin of pixels.
The work then becomes strangely delicate as well as technical. As if someone were carefully unrolling the scroll in virtual space, software traces the twisted layers of papyrus and digitally flattens them. The next trick is the real one. AI models look for minute ink traces in the scan that are invisible to the human eye.
It sounds simple, but it’s not. The ink used in ancient Greek manuscripts has nearly the same density as the papyrus itself, a problem that has plagued researchers for decades. The two look almost the same in CT scans. Computer scientist Brent Seales of the University of Kentucky spent years pursuing the issue, frequently running into obstacles that made the endeavor seem unrealistic. Then machine learning came into play.
Instead of teaching the system using ancient Greek letters, which would have required extensive linguistic training, researchers tried something easier and, looking back, more clever. The microscopic texture patterns connected to ink deposits were recognized by the algorithm. Greek was not “read” by it in the conventional sense. It recently discovered the ink. That was sufficient in some way.
After the AI combined those ink patterns, letter-like shapes started to emerge. Words came next. Luke Farritor, a young computer scientist, made an early discovery when he was able to decipher the first complete word concealed within a scroll: “purple.” Perhaps a small word. But it felt huge, standing there in digital ash.
Others soon joined the endeavor. A PhD candidate in Berlin named Youssef Nader advanced the models and started interpreting longer passages. Swiss robotics student Julian Schilliger improved the 3D mapping required to precisely flatten the scroll layers. They collaborated to win the $700,000 grand prize in the Vesuvius Challenge, a competition aimed at speeding up this kind of discovery.
Philodemus, a philosopher who resided in the villa where the scrolls were discovered, appears to be the source of the passages they were able to retrieve. He wrote about the good life, pleasure, beauty, and music—the kinds of subjects that seem strangely personal when you consider that the words survived a volcanic disaster.
That has a subtly eerie quality. These ideas were probably written while dinner was being prepared nearby in a cozy Roman library with a view of the Bay of Naples. The scene was frozen in time when the volcano erupted. These voices are now returning thanks to algorithms.
In other areas of historical research, the same technology is also making an appearance. In the last ten years, scholars have employed handwriting recognition systems to decipher faded letters, nineteenth-century diaries, and medieval manuscripts kept in European archives. Even skilled academics have long been frustrated by certain scripts, such as Beneventan. Strangely, computers don’t seem as intimidating.
The ramifications go beyond practicality. Historians can now examine thousands of manuscripts at once using AI transcription, comparing linguistic peculiarities or spelling patterns across enormous datasets. Millions of words could soon be used to answer questions that previously depended on a single, brittle document.
Nevertheless, the excitement is accompanied by a hint of skepticism. A lot of the time, machine learning acts like a mystery. Sometimes, even the engineers who create these systems acknowledge that they don’t fully comprehend how the models arrive at particular conclusions. Some historians may find this uncertainty unsettling in scholarship, where accuracy is crucial. However, it is hard to ignore the outcomes.
It’s difficult to ignore the peculiar reversal occurring when standing in front of those blackened scrolls in Naples, which are still fragile and sealed. Ancient manuscripts were endangered by technology for centuries due to fires, floods, and negligent restoration efforts. They are now being saved by technology.
We seem to be just getting started. There are still hundreds of unread Herculaneum scrolls. They might still contain entire philosophical works that are silently awaiting improved algorithms.
Additionally, another voice that has been forgotten might already be murmuring through the data somewhere in those layers of carbonized papyrus, just waiting for a machine to pick it up.
