The Herculaneum scroll sealed by Mount Vesuvius nearly two millennia ago has now been read across its surviving length without anyone physically opening it. Researchers combined high-resolution X-ray imaging, computational unwrapping and machine-learning assistance to recover about 22 columns of Greek from a papyrus that looks more like a lump of charcoal than a book. The result is not simply a better image of a famous artefact. It is evidence that one of the ancient world’s most inaccessible libraries can be approached as a collection of readable documents.

The achievement concerns PHerc. 1667, one of the carbonised papyri found at Herculaneum in the eighteenth century. Earlier attempts to open the roll damaged its outer layers and left a compact inner core. According to the Vesuvius Challenge announcement, the new reconstruction exposes roughly 1.4 metres of writing surface while leaving that core intact. Hundreds of unopened rolls and fragments remain, so a method that works repeatedly could change the surviving record of Greek and Roman thought.

The important word is repeatedly. One spectacular scroll makes a headline. A dependable pipeline for scanning, tracing, flattening, revealing and interpreting fragile papyrus could support years of scholarship. It also illustrates a useful truth about scientific AI: the system did not replace the physicists, computer scientists or papyrologists. It helped them see a surface that no unaided reader could reach.

The Herculaneum Scroll Crossed a Line Earlier Work Could Not

When LiveAIWire examined AI archaeology and digital dig sites in 2025, the Vesuvius Challenge had already produced remarkable partial readings. The June 2026 announcement is a genuine next step rather than a retelling of that story. The team reconstructed the writing across the preserved inner roll, joined it into a continuous reading order and submitted the image-supported text to specialist review.

That distinction matters because PHerc. 1667 was not an untouched, complete book. Nineteenth and twentieth-century efforts to open it destroyed or separated much of the exterior, leaving an inner cylinder about eight centimetres high from a roll once estimated at roughly 19 to 24 centimetres. The new work therefore recovers the surviving rolled core. It cannot restore sheets that were already lost.

The researchers’ technical preprint, which had not been peer reviewed at publication, describes 22 column-equivalents over about 860 square centimetres of preserved writing surface. The first three columns remain particularly fragmentary, with around 33 square centimetres not transcribed. Calling the result a full scroll reading is reasonable only with that boundary made clear: it is a full virtual unwrapping and scholarly reading of the image-supported surface that survives.

What This Means Beyond a Single Ancient Book

The Villa of the Papyri collection is exceptional because it preserves a substantial ancient library rather than isolated scraps quoted by later writers. The eruption of AD 79 carbonised the rolls, making conventional opening dangerous, but also helped preserve material that would otherwise have decayed, as the Vesuvius Challenge FAQ explains. Recovering even a fraction of the unread collection could add direct evidence to fields that often work from incomplete transmission and medieval copies.

For a non-specialist, the consequence is simple. Lost literature is no longer limited to objects that can safely be unfolded by hand. If a scroll can be scanned at sufficient quality and its layers can be mapped, a reader may eventually inspect a digital page while the original remains closed. That turns preservation and access from competing goals into parts of the same process.

It also connects with a broader change in cultural research. Projects using AI to digitise cultural heritage are moving beyond photographing visible surfaces. They increasingly reconstruct, classify and search material that is obscured, fragmented or too delicate to handle. The Herculaneum work is a particularly demanding example because the page itself first has to be found inside a crushed three-dimensional object.

How X-Rays Turned a Burnt Cylinder Into a Flat Page

The process begins with phase-contrast micro-computed tomography. Researchers scanned PHerc. 1667 at the European Synchrotron Radiation Facility’s BM18 beamline, producing a three-dimensional volume at a voxel size of 2.4 micrometres. The ESRF account of the project describes a surviving run of almost one and a half metres, with the writing distributed across roughly 20 columns when viewed as a continuous text.

A scan is not a photograph of a flat page. It is a dense map of layers that bend, touch, split and fold through the roll. Researchers must identify a papyrus surface inside that volume, build a mesh along it and transform the mesh into a two-dimensional view with as little distortion as possible. Where the automatic surface trace wandered or crossed layers, people corrected the geometry.

This virtual unwrapping problem has a longer history than the present competition. In 2016, researchers reported the non-destructive virtual unwrapping of the En-Gedi scroll, revealing a Hebrew text from Leviticus inside a damaged roll. Herculaneum presents an additional difficulty: its ink is mainly carbon and the papyrus has also been carbonised, so conventional absorption contrast offers little separation between writing and page.

The Machine Learning Model Did Not Read Greek

The phrase “AI read the scroll” is convenient, but it compresses several different jobs into one. The machine-learning models described by the researchers were trained to amplify image features associated with ink on a reconstructed papyrus surface. They were not trained on Greek letters, words or an optical character recognition task. They produced evidence that trained papyrologists could inspect, compare and transcribe.

This distinction is one of the strongest parts of the project. A model that predicts plausible language could silently complete damaged words from context and make invention look like recovery. A surface-conditioned ink model instead tries to expose marks supported by the scan. Scholars still decide whether a feature forms a letter, how uncertain readings should be recorded and what the resulting text means.

Independent experimental work supports the underlying idea that carbon ink can leave a detectable physical signature. A 2026 Scientific Reports study used machine learning with three-dimensional optical profilometry to distinguish written from unwritten papyrus through surface topography, while also finding that performance declined as resolution fell. Earlier research on Herculaneum ink recovery likewise showed that subtle surface morphology could make otherwise invisible writing more readable.

The human review is not ceremonial. It is the stage that separates an interesting image from defensible historical evidence. The same principle applies in other domains where AI flags material for specialists, including the recent AI paper checker that required authors and editors to confirm mistakes. High recall can be valuable, but the claim should remain no stronger than the evidence a qualified person can verify.

What the Recovered Text Actually Says

The papyrological reading of PHerc. 1667 indicates a work of Stoic ethics. The recovered passages discuss human nature, impulse, moral development and the reasoning needed for a good life. One important clue is the name Aristocreon, a nephew and student of the Stoic philosopher Chrysippus. That reference helps locate the work within the intellectual world of early Stoicism, although the author has not been securely identified.

The value of the text is not that it delivers one sensational sentence. It offers sustained argument from a philosophical tradition whose enormous ancient output survives only unevenly. Chrysippus reportedly wrote hundreds of works, yet none survives complete through ordinary manuscript transmission. Material connected with his circle can therefore clarify vocabulary, debates and lines of influence that later summaries cannot fully preserve.

This is why work on unread texts matters even when it does not reveal a famous lost title. A sequence of small recoveries can change how historians map schools, authors and concepts. Similar stakes arise when AI is used in efforts to preserve lost languages: the goal is not merely to accumulate symbols, but to retain the context that makes those symbols part of a human record.

Why “Complete” Still Needs a Careful Definition

The announcement invites two opposite mistakes. One is to dismiss the work because some letters remain uncertain and some exterior material was destroyed long ago. Ancient texts are routinely reconstructed from damaged evidence, with gaps and doubtful readings explicitly marked. The other mistake is to imagine a flawless digital facsimile in which every character is equally clear. The scan, model output and scholarly transcription contain different levels of confidence.

The researchers describe their models as visibility amplifiers rather than autonomous reading systems. That is a useful standard for reporting the result. A letter is strongest when the same shape is supported across raw scan features, reconstructed surface data, model output and papyrological context. A contextual guess alone is not equivalent to a visible stroke.

There is also a publication-status caveat. The full PHerc. 1667 methods paper is a preprint, so its workflow and interpretation can still change through formal review and subsequent scholarship. That does not erase the openly released images, meshes and transcriptions. It means readers should distinguish the observable dataset from every conclusion drawn from it.

Two Other Scrolls Tested Different Parts of the Method

The project did not rely on PHerc. 1667 alone. PHerc. Paris 4 contains thicker ink deposits that can be seen more directly at high scan resolution. Those deposits, estimated at about 10 to 20 micrometres thick, provided a validation case: researchers could compare visible surface evidence with the signals their ink-detection process highlighted.

A separate scroll, PHerc. 139, yielded a title identifying it as Philodemus’ On Gods, Book 8. A title can be disproportionately valuable because it links a physical object to an author, work and sequence. The result shows that the pipeline can produce both extended reading and targeted identification, two different scholarly benefits.

Together, the cases reduce the risk of treating one favourable result as universal proof. Paris 4 helps test whether the signal corresponds to real ink. PHerc. 139 demonstrates title recovery. PHerc. 1667 tests long-form surface reconstruction and reading. None guarantees that every unopened roll will respond equally well, but each constrains a different source of uncertainty.

Why Hundreds More Scrolls Will Not Be Read Overnight

The Vesuvius Challenge currently reports more than 600 unopened scrolls, 45 scanned items and just one fully read through the new pipeline. Its 2026 open-problems page describes surface extraction as still semi-automated and identifies automation, scale and generalisation across scrolls as central research problems.

Scanning time is one constraint. Geometry is another. Each crushed roll contains its own pattern of folds, cracks and touching layers, so a mesh that behaves well in one region may fail in the next. Ink visibility also varies with material, scan quality and the condition of the surface. Better models cannot recover detail that the imaging stage never captured.

Then comes interpretation. Ancient Greek papyrology is a specialised discipline, and a long text requires slow comparison of letter forms, gaps, syntax and historical context. More model output can create more work if it arrives without reliable confidence information or an interface that lets experts trace each reading back to the scan.

These limitations make the next stage less cinematic but more important. Researchers need tools that move from one-off demonstrations to documented, reproducible workflows. That pattern is familiar across AI-assisted archaeological discovery, where finding a candidate site or feature is only the start of verification, preservation and interpretation.

An Open Competition Became a Research Collaboration

The Vesuvius Challenge began as a prize competition, but the full-scroll effort depended on contributors sharing code, data and techniques. The successful pipeline combined work on imaging, segmentation, surface representation, ink detection and papyrology. Several people who entered as competitors later worked together, a useful sign that the problem rewarded cumulative progress more than a single secret model.

Open access is particularly important here because the objects are rare and physically inaccessible to most researchers. Releasing scans, derived surfaces, model outputs and transcriptions allows independent teams to challenge a reading or improve a component without rescanning the original. It also creates a record of how a conclusion was reached.

That transparency will matter as generated imagery becomes more convincing. A polished rendering of Greek letters is not evidence by itself. The credible chain runs back from transcription to model output, reconstructed mesh and measured X-ray volume. The project’s strongest legacy may be making that chain inspectable.

The Real Breakthrough Is a Repeatable Pipeline

PHerc. 1667 is already an extraordinary recovery, but its historical importance will depend on what happens next. If surface mapping becomes faster, scan protocols improve and ink models generalise, the result could turn a closed collection into a long-term digital excavation. New texts would arrive gradually, with uncertainties and revisions, rather than in a single dramatic reveal.

That is a healthier way to understand AI’s role. The technology did not resurrect an ancient author through prediction. It extended scientific instruments and human perception, helping specialists follow physical traces through layers of damaged material. The text remains an archaeological claim grounded in an object, not a plausible continuation generated from a language model.

The first fully read surviving core therefore marks both an achievement and a test. It proves that a carbonised roll can yield an extended, reviewable text without being opened. It does not prove that the remaining library will be easy. The hundreds of unread scrolls are waiting, but every one will still have to earn its way from scan to surface, from surface to ink and from ink to history.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.