AI reads a rolled Herculaneum scroll end-to-end, after decades of failed unwrapping

A sealed scroll from Herculaneum, destroyed by Mount Vesuvius nearly 2,000 years ago, has finally been read continuously end-to-end rather than as isolated words or patches. The breakthrough is centered on a previously unread rolled scroll, PHerc.1667, and the Vesuvius Challenge team says it was the first time the preserved text of a rolled Herculaneum scroll has been read end-to-end.

This matters because PHerc.1667 was not merely hard to interpret, it was physically damaged by earlier attempts to open it. According to the team, earlier efforts to open it by hand in the nineteenth century, and again in 1969 and in the 1980s, destroyed its outer layers, leaving only an 8 cm-high core of the original scroll. When standing upright originally, it measured between 19 and 24 cm in height. So the big promise here is not just “better imaging.” It is the ability to get systematic scroll-scale recovery from what’s left, without needing to preserve the risky physical unwrapping step.

Technically, the team’s success ties directly to a workflow upgrade and a higher resolution imaging approach. They used high-resolution phase-contrast X-ray microtomography performed at the European Synchrotron Radiation Facility in France. The paper described the imaging as an improved technique over prior methods used in the Vesuvius Challenge prize competition. In plain terms: rather than stopping at the kind of scan data that only supports partial, patch-based letter detection, they got enough fidelity, and paired it with software that could handle the geometry of a rolled, deformed object.

That “software that can handle the geometry” is doing a lot of heavy lifting. The team developed a workflow to scan the scrolls, detect ink on charred papyrus, virtually unroll the scrolls by modeling their deformed surfaces, and preserve those surfaces digitally. Those digital “surface models” then allow machine learning to identify letters across an entire scroll, not just in small isolated locations. The team summarized the key shift in their own words: “The key transition marked by the present work is therefore from exceptional local recovery to systematic scroll-scale recovery.” In other words, the win is scale, not just accuracy.

The project’s context is important for anyone tracking AI adoption beyond hype. Vesuvius Challenge is not the first time people tried to read ancient texts with ML and imaging. In 2023, researchers managed to decipher a few words among the char and ash that make up the bulk of the Herculaneum scrolls. Some of those prize-winning researchers later recovered more passages from one scroll, PHerc.Paris.4, and received the $700,000 grand prize from the Vesuvius Challenge contest in early 2024. Now, as the Thursday announcement and accompanying paper [PDF] described, those grand prize winners are part of the team that read a rolled scroll end-to-end in this latest experiment.

The Paris.4 thread also acts like a credibility check. The team says the new higher-resolution images taken for the latest experiment make the words on the scroll directly visible for the first time, removing the need to rely on algorithmic detection of individual words and phrases from CT scans. Most crucially for boards, funders, and anyone who cares about whether a model “hallucinates” plausible readings: the new scans of PHerc.Paris.4 perfectly matched what the grand prize team made out several years ago, providing independent confirmation that the prize went to the right team.

Beyond PHerc.1667, the team reports additional progress using the same workflow. They say PHerc.139 was determined to be a copy of book eight of epicurean philosopher Philodemus’ treatise On Gods. And they tied these findings to the broader objective of digitally unrolling the scrolls recovered from Herculaneum’s Villa of the Papyri, described in the piece as the world’s only surviving intact library from antiquity. In other words: the bigger target is an archive. If the workflow can be scaled across the hundreds of sealed scrolls recovered from the Villa of the Papyri, historians could expand what they can read, not just what they can guess.

Of course, scaling is not automatic. The team called out challenges that will shape whether this becomes routine or remains a breakthrough that only a few labs can repeat. They noted geometric challenges in surface prediction that can render an unrolled scan unreadable, and radiometric challenges that make ink identification difficult because ancient recipes were inconsistent. Those details matter to executives because they translate to real implementation risk. Imaging pipelines and ML models are only as scalable as their ability to handle messy inputs. Here, the “mess” is both physical deformation and material variability: the scroll surface is deformed, and the ink does not follow a single recipe.

Still, the direction of travel is unmistakable. The team believes their X-ray and machine learning workflow is ready to scale, and their framing is intentionally dramatic: “The thoughts of the ancient world, sealed in darkness for two millennia, are coming back into the light - a whole scroll at a time.” For decision-makers, the strategic stake is broader than archaeology. This is a template for any domain stuck behind physical constraints and incomplete data: use improved imaging to generate better signals, use an AI workflow designed for the geometry of the object, and move from “local wins” to “system-level recovery.” If that transition really holds across more scrolls, it could reshape what “readable” means for archives that have been written off as too damaged to decode.