Since 2023 the Vesuvius Challenge has paid out $1,868,000 to read the carbonized scrolls of Herculaneum without opening them. On 25 June 2026 the first scroll was read end to end — one out of forty-five scanned. Comparing it with the En-Gedi scroll of 2015 explains why the hard part was never the model, but the surface.
In 1750 a farmer digging a well on the slopes of Mount Vesuvius broke through the roof of a Roman villa buried by the eruption of AD 79. Inside was a library: papyrus rolls turned by heat into sticks of charcoal. For two and a half centuries, attempts to open them went the way such attempts go — some crumbled, some split, some were cut apart. More than six hundred are still closed.
On 25 June 2026 one of them — PHerc. 1667, catalogued as Scroll 4 — was read from beginning to end without being touched: roughly 1.4 metres of papyrus, about twenty-two columns of text. The work was done by a pipeline of machine learning models trained on phase-contrast microtomography data acquired on beamline BM18 at the ESRF synchrotron in Grenoble. The job of the artificial intelligence here is twofold, and worth naming precisely. First, inside the three-dimensional volume of the scroll, reconstruct the geometry of the rolled sheet — find where the papyrus surface runs through coils that are crushed and partly fused together. Second, on that virtually unrolled surface, detect traces of ink, which in X-ray data appear as minute density differences against surrounding charcoal. The output is not a translation or an interpretation. It is a flat image of a sheet nobody has opened. Reading it is still the papyrologists’ job.
The precedent nobody remembers: En-Gedi, 2015
The most useful historical comparison is not the Bourbon-era excavations but an episode from 2015. Brent Seales’s group at the University of Kentucky took a charred scroll found at En-Gedi, in Israel, and extracted its text — a passage from Leviticus — without unrolling it, using X-ray tomography. That scroll had been written with a metal-bearing ink: under X-rays it stood out. This is where the Herculaneum project technically begins, and the chronology matters, because it is the opposite of what people usually say. It is not that the Herculaneum method is now being exported to damaged archives around the world. It is that Herculaneum inherited a method invented elsewhere and had to make it far harder.
Harder for a mundane chemical reason. The ink on the Herculaneum papyri is carbon-based, like the carbonized papyrus it sits on. Dark material on dark material, nearly identical density. There is no contrast to exploit, and for years the problem looked physically unsolvable. What machine learning contributed was not new contrast: it was the ability to recognise a statistical signature — a texture pattern in the volume — that the human eye cannot separate from noise.
How you win $700,000 with four passages of text
The Vesuvius Challenge launched on 15 March 2023, backed by Nat Friedman, Daniel Gross and Brent Seales, with over a million dollars in initial prizes and scrolls from the Institut de France scanned at the Diamond Light Source in Oxford. The structure is an open contest: data released under a Creative Commons licence, code on GitHub, prizes tied to milestones.
On 5 February 2024 the 2023 Grand Prize, $700,000, went to Youssef Nader, an Egyptian PhD student in biorobotics in Berlin, Luke Farritor, a 21-year-old from Nebraska and SpaceX intern, and Julian Schilliger, a robotics student in Zurich. The criterion was measurable and modest: four passages of 140 characters each with at least 85% legible text. They delivered over two thousand characters, eleven columns more than required.
It is worth pausing on who won. Not a papyrology department, not a heritage-science lab: three young people with computer vision and robotics skills, working on public data. This is the prize-based model pushed to its limit, and it worked in a domain where the traditional academic community had neither the tools nor the incentives to move that fast.
But the papyrologists were not spectators
Here a common narrative needs correcting — the idea that philologists passively received results from the engineers. In January 2024 a panel of papyrologists reviewed fifteen columns in anonymised form, not knowing which team had produced them, in order to judge whether the text was genuinely legible. They were the ones who recognised the content: an Epicurean text on music, food and pleasure, mentioning someone called Xenophantus. The legibility judgement — the prize criterion itself — was a philological judgement.
And one detail is worth more than most press statements. On PHerc. Paris 4, the Grand Prize scroll, later higher-resolution scans show the ink directly visible in the 3D data. Compared with the 2023 reading, it matches one to one. That is as close to independent verification as this field can produce: a model trained to guess where ink was had guessed correctly.
The scoreboard, with the real numbers
This is where the triumphal story meets the accounting. Prizes actually paid out since 2023: $1,868,000, spread across about ten milestones — $100,000 for Kaggle ink detection in July 2023, $60,000 for first letters in October 2023, $60,000 for first automated segmentation in December 2024, $60,000 to Marcel Roth and Micha Nowak in May 2025 for finding the end-title of a still-sealed scroll, $200,000 for Kaggle surface detection in March 2026.
- Scrolls and fragments scanned: 45.
- Scrolls read in full: 1.
- Scrolls where ink has emerged: 9 out of 45, and not always legibly.
The 2024 Grand Prize, announced with a target of 90% of four scrolls, does not appear among the awards paid. The roadmap has slipped by roughly two years, and the goal is now reframed as a $1,000,000 Grand Prize 2027: fully unwrap one of thirteen sealed scrolls, deadline 25 June 2027. The total open pool is $2,140,000, including $500,000 in First Letters prizes ($50,000 per scroll) and $590,000 a year in Progress Prizes, $20,000 monthly to the best contribution.
Technically, the project is more candid than its headlines: surface tracing is described as semi-automatic, and generalising ink detection from one scroll to another remains an open problem. That is the part the public struggles to see. A model that reads PHerc. Paris 4 well does not automatically read the scroll next to it, because each scroll has its own history of heat, deformation and damage. PHerc. 1667 had already been attacked in the nineteenth century, again in 1969, and again in the 1980s: what survives is the inner core, about 8 cm of an original height of 19–24 cm. The bottleneck is not ink recognition. It is geometry.
What it said, and what we cannot claim
PHerc. 1667 contains a philosophical treatise on ethics with Stoic indications; the last surviving column names Aristocreon, nephew and pupil of Chrysippus, and from language, themes and that name a second-century BC date has been proposed. No author’s name appears in the text: the framing is an inference and should be treated as one. On another scroll, PHerc. 139, title and attribution have been recovered — Philodemus, On the Gods, book 8 — and in February 2025 the Bodleian Library in Oxford announced the first images from inside one of its Herculaneum scrolls. The 2026 full reading currently exists as an arXiv preprint; the scans were made in collaboration with the National Library of Naples.
One figure changes the scale of the problem. Most of the money comes from a handful of private donors: Nat Friedman has put in $2,250,000, the Musk Foundation $2,084,000, followed by Alex Gerko, Joseph Jacks and Daniel Gross. Which means the speed at which Europe re-reads the only ancient library it has left depends, today, on a few people in California finding it interesting. It has been effective. It is not a cultural policy.