When AI Read Two Thousand Year Old Scrolls

Image on Unsplash

In 79 AD, Mount Vesuvius had erupted and buried Herculaneum under volcanic ash. As a consequence, hundreds of papyrus scrolls in a wealthy Roman's library were preserved in a state of near-perfect carbonization because they were not exposed to oxygen or decay (Conversation, 2024). For two millennia, they sat underground, unreadable and untouched, and then AI machines learned how to see inside them without destroying them.

The breakthrough happened when computer scientist Brent Seales partnered with tech investors to create the Vesuvius Challenge, which offered 700,000 dollars to anyone who could read the inside of a rolled-up ancient scroll using only non-invasive imaging and artificial intelligence (Smithsonian, 2025). The challenge was that opening the scrolls physically would destroy them because they are carbonized to the point where any physical manipulation causes them to disintegrate.

What Seales and his team did was take CT scans of the scrolls, which create high-resolution cross-sectional images that reveal the layers of papyrus wound inside the roll. Then, they trained machine learning algorithms to recognize what ink looks like versus what blank papyrus looks like. The AI learned this distinction from fragments of opened scrolls and from the written records that historians already have, and then the algorithm applied that knowledge to the sealed scrolls by analyzing the CT scan data layer by layer (Smithsonian, 2025).

By 2023, the team had successfully read one scroll without opening it at all, and then they set a goal to read 90 percent of four scrolls by 2024. 2000 years of written information that was thought to be permanently lost has suddenly become accessible (Conversation, 2024).

The texts that have been recovered so far include business letters and philosophical writings and receipts, as well as personal correspondence from people who lived during the eruption. The scrolls were transformed from abstract historical artifacts into actual windows into how people in 79 AD lived and thought about their daily lives.

This was a very useful blend of technology and traditional scholarship that neither side could have accomplished alone. The machines can process the images faster than humans could ever do manually, but humans are still needed to interpret what the machine learning algorithms have revealed.

DeepMind also created a system called Ithaca that works with ancient Greek inscriptions, which it trained on 78,608 inscriptions spanning from the seventh century BC to the fifth century AD. The system can not only restore fragmented text but also predict when and where the inscription was created based on linguistic patterns that the machine learning model learned (Nutanix, 2025).

Cambridge classics professor Mary Beard described the traditional process of identifying ancient inscriptions as essentially guesswork or following hunches. But now, Ithaca can scan its database of thousands of inscriptions and surface relevant parallels almost instantly. That accelerates the research process and allows scholars to test hypotheses that would have taken months to test manually (YouTube, 2025).

The machine learning models are now attempting to decipher completely unknown ancient writing systems like Linear A and the Cypro-Minoan script, which have never been successfully translated despite thousands of years of trying.

The approach these systems use is mathematical rather than linguistic in the traditional sense, because they look for patterns that suggest one sign maps to another sign. Then, they test whether those mappings produce words that appear frequently enough in comparable texts to suggest actual meaning rather than random associations.

This causes a shift in how archaeology and historical scholarship work, because previously the constraint was human time and human expertise. Now, the constraint is data quality and computational resources.

There are still logistical challenges because producing the high-resolution CT scans requires access to particle accelerator facilities that are limited and expensive. Besides, most of the unopened scrolls from Herculaneum are stored in Naples and transporting them safely poses both logistical and financial hurdles.

But the point is that thousands of texts are no longer considered permanently lost, which changes how historians and archaeologists think about what is recoverable from the past.

It also raises a question about what else has been considered lost that might be recovered, given that we had the right combination of technology and expertise applied to it.