For nearly two thousand years, the words inside a carbonized Herculaneum scroll remained physically present and practically unreachable. Opening the brittle roll risked destroying the evidence.
A research team has now made the preserved writing surface of PHerc. 1667 readable without physically opening it. High-resolution synchrotron X-rays captured the scroll’s interior. Software reconstructed and flattened the wound sheet. Machine-learning models amplified subtle traces of carbon ink. Then eight papyrologists inspected the evidence, transcribed the Greek and resolved readings independently or by consensus.
The breakthrough is not that an AI translated an ancient book. It is that machines extended what people could perceive—and an open scientific community turned that new visibility into scholarship.
Research references: [1]
From sealed object to inspectable surface
The scroll was scanned on the BM18 beamline at the European Synchrotron Radiation Facility in Grenoble using phase-contrast X-ray microtomography. The scan produced a three-dimensional volume of tightly wound, crushed papyrus layers.
That volume was not a stack of ready-made pages. Researchers still had to trace a single sheet through the scan, correct places where geometry jumped or merged, and flatten the resulting mesh into a two-dimensional surface. The finished reconstruction covered 31 wraps and 1,231 square centimeters of papyrus. Even with the team’s most efficient semi-automated workflow, the paper reports roughly 25 hours of manual annotation per wrap.
Machine learning addressed a different problem: making faint ink evidence easier to see. Because both ink and papyrus were carbonized, their signals could be difficult to separate. The models acted as visibility amplifiers. They were not trained on letters, words, transcriptions, OCR targets or vocabulary, and no OCR or language model generated the reading.
Human experts remained responsible for the text. Eight qualified papyrologists compared flattened renderings, ink-enhanced outputs and the geometry of the underlying layer. A reading was accepted only when the surface fit the papyrus in the scan, letterforms remained consistent across repeated renderings or inference passes, and experts endorsed the transcription independently or reached consensus.
Research references: [1]
Twenty-two columns returned to view
The reconstructed surface contains about 22 columns or column-equivalents across roughly 860 square centimeters of preserved writing area. The first three surviving columns include 33 square centimeters too fragmentary for secure interpretation; the transcription marks physically lost text and uncertain letters using standard papyrological conventions.
That boundary is central to the claim. PHerc. 1667 is the first Herculaneum papyrus whose preserved, still-rolled writing surface has been digitally unwrapped and read continuously for extended scholarly study. “Complete” does not mean every word once written on the original roll survived.
Earlier attempts to open the object had destroyed outer layers. The remaining core is about eight centimeters high, compared with an estimated original height of 19 to 24 centimeters. The achievement is a complete reconstruction and reading of the defined preserved surface—not recovery of the lost exterior.
The surviving text appears to be an ethical treatise concerned with human nature, impulse and moral development. Its title and author remain unknown. The final preserved column appears to name Aristocreon, a nephew and student of the Stoic Chrysippus; together with the language and themes, that points toward a Stoic context. The paper dates the manuscript to the second century BC while not excluding the very late third century BC.
Technology recovered evidence. Papyrology supplied the reading and interpretation.
Three scrolls, three kinds of proof
The paper separates three results rather than treating every scroll as the same test.
PHerc. 1667 demonstrates end-to-end virtual unwrapping and reading across the preserved scroll surface.
PHerc. Paris 4 provides a physical validation case. Under an optimized high-resolution scan, ink-bearing deposits became directly visible in the three-dimensional volume. Researchers segmented those deposits and projected them onto the flattened surface, where they matched text recovered in the Vesuvius Challenge’s 2023 work. The authors carefully limit this validation to at least one scan regime and preservation state.
PHerc. 139 provides a title-identification result. From its reconstructed title region, scholars read an author, work and book number: Philodemus, *On Gods*, Book 8. That means a sealed roll may yield the ancient equivalent of library metadata even before its continuous prose is recovered.
Together, those cases show three different capabilities: follow a preserved rolled text end to end, test recovered ink against directly visible deposits, and identify a closed work from its title region.
Research references: [1]
The community became part of the method
The technical work grew through the Vesuvius Challenge, a donation-funded open competition built on virtual-unwrapping research pioneered at EduceLab. The project released data, code, models and tractable subproblems so people outside the original institutions could participate.
According to the project, many researchers on the team first arrived as contestants, won prizes for their contributions and were later recruited. Specialized equipment, conservation access and scholarly expertise remained essential. But open artifacts widened who could help solve the computational pieces.
The paper’s data-availability section points to public volumes, reconstructed surfaces, renderings and transcription artifacts. It also links a public code commit sufficient for the principal mesh-tracing, flattening, rendering and ink-inference analyses.
That is a second kind of breakthrough: the research did not only reveal a hidden text. It created pathways for engineers, students and independent contributors to become part of the team.
One scroll is the beginning
This result does not make every sealed Herculaneum papyrus automatically readable. Scan quality varies within and across objects. Compressed layers can blur together. Surface tracing can jump between sheets or leave holes. Ink models that work on one scroll may plateau or reveal little convincing signal on another.
The Vesuvius Challenge’s October 14 technical roadmap says the next goal is reliability at collection scale. Human correction is still necessary, and the pipeline must become better at distinguishing “no ink” from “no ink recovered yet.”
Those limits sharpen the human significance of the result. For the first time, a preserved rolled Herculaneum text has moved from sealed object to sustained scholarly reading without the object being opened. The path combined rare imaging infrastructure, painstaking geometry, machine-learning assistance, open-source engineering and expert judgment.
The scroll stayed closed. Its conversation with humanity reopened.
*Disclosure: Next for Humanity independently reviewed the current arXiv record, full preprint, project milestone and October 14 technical roadmap. The paper remains arXiv version 1, submitted June 27, 2026, and is labeled a preprint rather than a peer-reviewed publication. The authors disclose Vesuvius Challenge support, earlier Mellon Foundation support for EduceLab groundwork, no competing interests and limited large-language-model assistance in manuscript preparation; they state that human authors reviewed and approved all scientific claims and final wording. This article was prepared with AI assistance and human editorial review.*
Research file / primary sources
Evidence behind this analysis
Research file updated October 25, 2026. These primary and official sources were accessed on the dates shown. Notes explain how each source informed this analysis; inclusion does not imply endorsement.
- Complete virtual unwrapping and reading of a rolled Herculaneum papyrus (opens in a new tab)
arXiv · Published June 27, 2026 · Accessed October 25, 2026
Primary full preprint covering methods, evidence boundaries, transcription, review, funding, disclosures, data and code.
- arXiv record for 2606.29085 (opens in a new tab)
arXiv · Published June 27, 2026 · Accessed October 25, 2026
Primary version and bibliographic record; v1 remains current as of access.
- We read an entire scroll — without ever opening it (opens in a new tab)
Vesuvius Challenge · Published June 25, 2026 · Accessed October 25, 2026
Primary project milestone covering the physical history, open-science organization and community pathway.
- Open Problems: Why Reading Every Herculaneum Scroll Is Still a Challenge (opens in a new tab)
Vesuvius Challenge · Published October 14, 2026 · Accessed October 25, 2026
Primary current technical roadmap describing limits, human correction and cross-scroll generalization.
This piece is original analysis from Next for Humanity. We separate editorial judgment from commercial relationships, update material errors, and disclose owned products where they appear. Read our editorial standards.