Jump to content

Reading the sealed Herculaneum scrolls with AI

From IdeaWazaWiki

How Open Data, Cash Prizes, and Machine Learning Began Reading Scrolls Burned Shut by Vesuvius

The Herculaneum scrolls are a library of ancient papyrus rolls that the eruption of Mount Vesuvius turned into lumps of carbon in 79 AD. For almost three centuries nobody could read the sealed ones without destroying them. Since 2023 a mix of particle accelerator scans, machine learning, and open cash prizes has started to read them while they stay rolled up. This article is about futurism and about moving research from scarcity toward abundance: the story shows that the scarce input was never the scrolls or the scanner, but skilled attention, and that publishing the data and paying for verified open results can turn a handful of experts into a worldwide crowd of contributors.

The problem that stumped scholars for 275 years

Workers found the scrolls in the 18th century in a villa at Herculaneum, near Pompeii. Hundreds are now kept in a library in Naples. Starting in the 1750s, people tried to open them by force. The first serious attempt was Antonio Piaggio's mechanical unrolling device, a frame of weights and silk threads. Later attempts used chemicals, gas, and slicing. These methods often cracked or shattered a scroll. Some pieces survived well enough to be read, and many others were damaged or destroyed outright.

The physics made a gentler path hard as well. A normal X-ray scan shows where materials differ in density. The scrolls were written in carbon ink on papyrus that the eruption had also turned into carbon, so the ink and the page are made largely of the same element and give very little contrast in a raw scan.

Virtual unwrapping before the prize

The first proof that a burned scroll could be read without opening it came from a different object. In 2016 a University of Kentucky team published a study in Science Advances on the scroll from En-Gedi, a charred Hebrew scroll found in an ancient synagogue in Israel. Their software pipeline, which they called virtual unwrapping, had three steps: find the writing surface inside a 3D scan (segmentation), assign the scan values on that surface to an image (texturing), and flatten the curved surface into a flat page. The result revealed five complete wraps of text from the book of Leviticus.

En-Gedi had ink that showed up in the scan. Herculaneum ink mostly does not, so the same method needed a new way to find ink that the human eye cannot see in the data.

The prize model

On March 15, 2023, the Vesuvius Challenge launched. Scrolls from the Institut de France had been scanned at the Diamond Light Source, a particle accelerator near Oxford. The organizers released the high resolution CT scans to the public and offered more than one million US dollars in prizes put forward by private donors. Winning code was published as open source, so each winning method became a starting point for everyone else.

The trick that unlocked the ink is clever. Some fragments that broke off during the old unrolling attempts show their writing on the surface. Those fragments can be photographed, often in infrared, and also CT scanned. That gives a training pair: the scan as input and the visible ink as the answer. A model learns from the fragments and is then applied to the sealed scrolls.

Progress came fast. In 2023 a college student was the first person to read a whole word from inside a sealed scroll, the Greek word for purple. In December 2023 a team of three students reached the Grand Prize goal: four passages of 140 characters each with at least 85 percent of characters recoverable. Their submission also included 11 more columns of text, more than 2,000 characters in total. The organizers wrote that most of them had given the goal less than a 30 percent chance of success when they announced it. The team won 700,000 US dollars, and three runner up teams won 50,000 US dollars each. Papyrologists reviewed anonymized entries and a technical team reproduced the code, a guard against models that invent letters.

What the first scroll says about abundance

The text recovered in 2023 covers only about 5 percent of one scroll, but it was a work nobody had seen before. Its subject is pleasure, the central good in Epicurean philosophy, and the author is likely Philodemus, the philosopher who seems to have worked in the villa library. One passage reads, in the organizers' translation, "as too 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."

That line is a fitting first message from a library that was itself locked in scarcity for two thousand years. An ancient writer questioned the idea that scarcity makes things more valuable, and the means of reading him came from making the data abundant.

The bottleneck moved from ink to labor

Reading ink was only half the job. Someone had to trace the papyrus surface through the crushed spiral of the scroll first, and in 2023 that work was mostly manual. The organizers reported that it cost more than 100 US dollars per square centimeter of manual labor to produce the readable text, and that at this price, tracing every scroll could cost hundreds of millions or even billions of dollars.

So the next prizes targeted automation. The 2024 goal was to read 90 percent of four scanned scrolls, and it went unclaimed. A separate prize asked for an automated method that could reproduce the 2023 tracing with no more than four hours of human input. Two strong entries came close and each received 30,000 US dollars, but an independent review found that neither recovered as much text as the 2023 result. In early 2025 the organizers also said they still had not recovered text conclusively from three of their other scrolls.

Then results began to spread. In May 2025 two graduate students won a 60,000 US dollar prize for reading the title inside PHerc. 172, a sealed scroll held at the Bodleian Libraries in Oxford. It identified the work as Philodemus, On Vices, probably Book 1. A member of the organizers' research team found the same title independently, a useful cross check. In 2026 the project reported that another sealed scroll, PHerc. 1667, had been virtually unwrapped and read.

Agents working beside people

One 2026 detail shows how AI agents can help human researchers without replacing them. The organizers reported that an autonomous agent swarm, adapted from an open source automated research project, tested model settings continuously and found a configuration that nearly doubled a validation score on PHerc. 1667 while training only on data from a different scroll. The score was computed against approximate labels, so it is a signal, not a final reading. Papyrologists still decide what the text says.

Honest limits

The project itself says it cannot claim final victory. Its 2026 open problems page lists the hard parts plainly:

  • Ink models that work on one scroll may fail on another, and the reasons are sometimes unknown.
  • Heavily compressed regions scatter the X-ray beam, and whole scrolls cannot yet be scanned at the highest resolution in practical production.
  • The team cannot always tell the difference between a region with no ink and a region where the ink has not been recovered yet.
  • Automated tracing still needs human correction, and the scan data for a large scroll can reach dozens of terabytes.

The current rules reward tools that are released early, documented, and used by others, to discourage secret solutions. Some scan data is published under a noncommercial license (CC BY-NC 4.0), so it is open for learning and research rather than fully public domain. The next target is a one million US dollar Grand Prize planned for June 2027 for showing that the full process can work reliably and at scale.

Why it matters beyond the classics

The organizers estimate that the scrolls in Naples alone hold more than 16 megabytes of text, and parts of the villa remain unexcavated, so more scrolls may still be buried. The wider lesson reaches past ancient philosophy. Many research problems are stuck less by a lack of ideas than by a lack of people who can see the data. The Vesuvius model has four parts that other fields can copy: release the raw data, define a measurable goal, require winning methods to be open, and have independent experts check the results.

  • Why did releasing the scan data to the public speed up progress more than giving it to a single expert lab would have done?
  • Which unsolved research problem in medicine, aging, or materials could be organized like the Vesuvius Challenge, and what would its measurable goal be?
  • How should a prize competition protect itself against machine learning models that produce letters or patterns that are not really there?
  • What does the recovered passage about scarce and abundant food suggest about how people judge value, and does modern economics agree?
  • Should cultural data such as scans of ancient objects be fully public domain, or is a noncommercial license a fair middle path?
  • How can AI agents that tune models around the clock be paired with human experts so that people keep final judgment over what counts as a discovery?
  • What is lost and what is gained when a field depends on wealthy private donors to fund its biggest breakthroughs?

Sources

See also