How AI Helps Read Ancient Texts: Ithaca and the Scroll Prize

How the Ithaca network and the Vesuvius Challenge used AI on damaged ancient texts, what the figures measure, and what they leave open.

Guide: AI BasicsLast verified: October 20265 min read

Artificial intelligence now works on historical sources themselves, not only on the history of computing. Two documented cases show how: a neural network named Ithaca, described in Nature on March 9, 2022, that helps historians restore, place, and date damaged ancient Greek inscriptions, and the Vesuvius Challenge, in which machine learning helped read a carbonized Roman scroll that cannot be unrolled. In both, the software proposes and people with training in the sources decide.

Key takeaways

  • Ithaca alone restored damaged text with 62 percent accuracy in its paper; historians using it rose from 25 percent to 72 percent.
  • The same paper reported 71 percent accuracy for geographic attribution and dates within 30 years of the ground-truth ranges.
  • The Vesuvius Challenge announced its Grand Prize of USD 700,000 on February 5, 2024, after a team read more than 2,000 characters from a scroll that was never opened.
  • Both projects measure success against known answers or set thresholds, which is why the results are specific and limited.

The problem: sources that are broken or sealed

Much of what is known about the ancient world comes from objects that did not survive intact. Inscriptions on stone lose letters to weathering and breakage. Papyrus rolls can be so fragile or so fused by heat that opening them destroys them. A historian working on such a source has two jobs: recover as much text as possible, and say where and when it was made. Both rest on pattern knowledge, such as which words, formulas, and letter shapes appeared in which place and period. That is the kind of knowledge a trained model can store at scale, and it explains why this field attracted machine learning researchers.

Ithaca: restoring, placing, and dating Greek inscriptions

The Nature paper "Restoring and attributing ancient texts using deep neural networks," by Yannis Assael, Thea Sommerschield, and colleagues from DeepMind, Oxford, Ca' Foscari University of Venice, and the Athens University of Economics and Business, reported three results for Ithaca. On restoring damaged text, the network alone reached 62 percent accuracy. Geographical attribution reached 71 percent, and the chronological dates it gave fell within 30 years of the ground-truth ranges. The finding with the most practical weight is the human one: when historians used Ithaca, their own restoration accuracy rose from 25 percent to 72 percent.

These figures describe performance on the paper's own test material, and the paper's full methods are the place to check how that material was chosen. The numbers support a modest reading: the network is useful as an assistant to a specialist, and the specialist's judgment is still the step that decides what is published. A restoration is a hypothesis about lost letters, and for a truly lost text no answer key exists.

Vesuvius Challenge: reading a scroll without opening it

The eruption of Mount Vesuvius in AD 79 buried a library at the Villa of the Papyri near Herculaneum under about 20 meters (about 66 feet) of volcanic material. Heat carbonized the papyrus scrolls, which preserved them and made them brittle. The Vesuvius Challenge, an open competition with monthly prizes, set out to read them virtually. The workflow had three stages: X-ray computed tomography scans, taken at the Diamond Light Source particle accelerator near Oxford; segmentation, which traces and flattens the rolled layers on screen; and ink detection, where machine learning models mark which regions of the flattened surface carry ink.

On February 5, 2024, the organizers announced the Grand Prize of USD 700,000 for a team of Youssef Nader, Luke Farritor, and Julian Schilliger. Their submission met the contest criterion of four 140-character passages with at least 85 percent of characters recoverable, and it included 11 further columns, more than 2,000 characters in total. The organizers describe the text as philosophical and about pleasure, likely by an Epicurean author such as Philodemus, and say roughly five percent of the scroll had been read. Hundreds of other scrolls remain, so the method has been shown on a fraction of the material.

The two projects side by side

QuestionIthacaVesuvius Challenge
Source materialAncient Greek inscriptions, damaged but visibleCarbonized papyrus scrolls, sealed
What the software doesSuggests restorations, places, and datesMarks ink in flattened scan images
Headline result62 percent alone, 72 percent with historiansMore than 2,000 characters read
Who checks the outputEpigraphers, who read the suggestionsPapyrologists, against the contest's recovery threshold
Main cautionAccuracy is measured on known texts, not lost onesAbout five percent of one scroll read so far

What these cases do not show

Neither project shows that software understands history. Ithaca finds patterns in inscriptions it has seen; the contest's models find ink-like texture in scans. A confident output can still be wrong, and a wrong restoration, once published, can steer later scholars. The dating figure also deserves careful reading: "within 30 years" is a measure of distance from the known range, not a promise that each date is that close.

The cases also depend on human work that the headlines skip. Training material had to be assembled in usable form, and the Vesuvius teams had to build the segmentation tools before ink detection could start. The model is the last link in a chain of preparation.

Questions to ask about any AI-and-history claim

  • What was the accuracy measured against, and could the test material have been seen in training?
  • Does the headline number describe the software alone or the software with experts?
  • Are the outputs offered as suggestions, with the uncertainty visible?
  • Who with training in the source type reviewed the result?
  • How much of the whole collection has actually been processed?

FAQ

Does AI write history?

No. In these projects it assists with reading or restoring primary sources, and historians interpret the results. Historical narrative and argument remain human work.

Why does the Ithaca result of 72 percent matter more than 62 percent?

The 62 percent figure is the software alone. The 72 percent figure is what historians achieved with the software, up from 25 percent without it, which is closer to how the tool would be used in practice.

Were the Herculaneum scrolls physically opened?

The Vesuvius Challenge describes reading them without unrolling them, using scans and software. The winning team worked on a scroll held by the Institut de France.

Is the history of AI itself covered here?

No. This article covers AI used on historical material. The development of AI as a field is traced in separate timeline and background articles.

Last verified: October 2026

Sources

  • Assael Y, Sommerschield T, Shillingford B et al., Restoring and attributing ancient texts using deep neural networks, Nature 603, 280-283, March 9, 2022
  • Vesuvius Challenge, 2023 Grand Prize announcement and project overview, Vesuvius Challenge, February 5, 2024

New to artificial intelligence?

Start with the basics, then follow the guides to tools, rules, and effects on society.

Start with AI Basics