For nearly 2,000 years, hundreds of papyrus scrolls lay buried under volcanic mud, carbonized into fragile lumps of charcoal. These scrolls, preserved by the eruption of Mount Vesuvius in 79 AD, were considered unreadable. Attempting to open them would turn the ancient paper to dust. However, a major breakthrough occurred recently when a team of students used artificial intelligence to read the text inside the scroll without ever physically touching it.
The scrolls originate from the “Villa of the Papyri” in Herculaneum, a luxury Roman estate believed to have belonged to the father-in-law of Julius Caesar. When Vesuvius erupted, the heat carbonized the scrolls instantly. Unlike Pompeii, which was buried in ash, Herculaneum was buried in deep mud, which preserved the organic material but fused the scroll layers together.
Since their discovery in the 1750s, historians have tried everything to open them. Early attempts involved mechanical machines that often destroyed the outer layers. More recently, scientists realized they needed a way to see inside the scrolls without unrolling them.
The primary challenge was not just the fragility of the scrolls. It was the chemistry. Most ancient documents were written with iron-based inks, which show up clearly on X-rays. However, the Herculaneum scrolls were written with carbon-based ink (charcoal and gum).
Because the papyrus itself was turned into carbon by the volcano, the ink and the paper have the exact same density. On a standard X-ray or CT scan, the letters are invisible. This rendered the library a “black box” to archaeologists for decades.
In 2023, computer scientists Nat Friedman, Daniel Gross, and Professor Brent Seales launched the Vesuvius Challenge. They released high-resolution CT scans of the scrolls to the public and offered over $1 million in prizes to anyone who could create code to identify the ink.
Professor Seales, from the University of Kentucky, had already spent two years developing “virtual unwrapping” software. This software takes the 3D data from a CT scan and flattens the rolled-up layers into 2D sheets. The missing piece was detecting the ink on those sheets.
In early 2024, the organizers announced that a team of three students had won the $700,000 Grand Prize. They successfully deciphered more than 2,000 characters from a single scroll. The winning team included:
The team did not use magic or guessing. They used a specific type of machine learning called a ResNet (Residual Neural Network) and a UNet architecture. Here is how they accomplished the task:
The result was a clear image of Greek letters emerging from the black background of the CT scan.
The text revealed by the AI is a philosophical work. Scholars have identified the author as Philodemus of Gadara, an Epicurean philosopher who lived in the first century BC. He was essentially the “resident philosopher” of the villa.
The recovered text does not discuss wars or politics. Instead, it focuses on the senses and pleasure. Philodemus writes about music, food, and how to enjoy life.
The AI revealed several specific arguments Philodemus makes:
This discovery confirms that the library contains unique texts that do not exist anywhere else in the world.
The success of the Vesuvius Challenge is one of the most significant archaeological events of the 21st century. The Villa of the Papyri is the only intact library surviving from the Greco-Roman world.
Currently, historians have only excavated a small portion of the villa. Archaeologists believe hundreds or even thousands more scrolls remain buried in unexcavated sections of the house.
This technology opens the door to finding “lost” masterpieces. Historians hope to find:
With the proof of concept complete, the project is moving to the next phase. The goal for 2024 and beyond is not just to read a few paragraphs, but to read entire scrolls from beginning to end.
The challenge organizers have set a new goal to read 90% of the four scrolls currently scanned. They are also working to automate the “segmentation” process (tracing the layers of papyrus), which is currently the most time-consuming part of the workflow. As the AI improves, the cost of reading a scroll will drop, allowing museums around the world to unlock the secrets inside their own unreadable artifacts.
What language are the scrolls written in? The scrolls deciphered so far are written in Ancient Greek. However, scholars believe there is a “Latin section” of the library yet to be fully explored, which could contain different types of literature.
Can the scrolls be physically opened now? No. Physically opening them still carries a high risk of destroying them. The virtual unwrapping method is the only safe way to access the text.
Who owns the scrolls? The scrolls are held by the National Library of Naples in Italy and the Institut de France in Paris. The Vesuvius Challenge works in partnership with these institutions to scan the artifacts.
How much of the library has been read? Only about 5% of the text from one scroll has been fully deciphered as of the latest major announcement. There are approximately 800 scrolls that have been excavated, with potentially thousands more still underground.
Is the AI guessing the words? No. The AI is not a language model like ChatGPT that predicts the next word. It is a computer vision model. It detects the presence of ink pixel by pixel. If the AI shows a letter “Alpha,” it is because the physical ink is there on the page, not because the computer thinks it should be there.