The Herculaneum Scrolls — Reading the Dead with AI

In August 2023, a 21-year-old undergraduate at the University of Nebraska detected the Greek word πορφύρας — "purple" — inside a lump of carbonized papyrus that had been illegible for 1,944 years. The scroll was never unrolled. It was read by X-rays and a neural network trained to spot a 1–2 Hounsfield-unit difference between ink and the carbon it was burned into.

The library it came from is the only intact one to survive from the classical world. Roughly 1,800 scrolls were dug out of the Villa of the Papyri at Herculaneum starting in 1752 — buried at 300°C under Vesuvius's pyroclastic surge in 79 AD. For 271 years the still-rolled ones were considered unreadable. They were wrong, but only by a few decades.

How the read works

The carbon ink — soot, gum arabic, water — sits on carbonized papyrus. Visually: black on black. In a CT scan at 4–8 μm resolution, the ink has slightly higher X-ray attenuation than the substrate. The difference is roughly 1–2 Hounsfield units, smaller than the scanner's noise floor. The signal exists; it just doesn't survive a single voxel.

The pipeline that won the prize money:

Scanning at Diamond Light Source (UK) and the European Synchrotron (Grenoble). Hours of beamtime per scroll.
Segmentation — tracing each papyrus surface through the 3D volume, a geometry problem closer to cave-mapping than image processing. The scroll is wound thousands of times; the surface twists, cracks, jumps layers.
Ink detection — ensemble ML models trained on physically unrolled reference fragments, learning the texture signature ink leaves even when its density signal alone is buried in noise.
Papyrology — humans reading the rendered text and arguing about it.

Stage 3 is where the prize-winners broke through. Stage 2 is what still gates full-scroll recovery.

Timeline

March 2023 — Brent Seales (University of Kentucky) and Silicon Valley funders launch the Vesuvius Challenge. $1.7M in prizes.
August 2023 — Luke Farritor wins the $40,000 First Letters Prize for πορφύρας.
October 2023 — Youssef Nader independently confirms the method on a separate fragment.
February 2024 — Nader, Farritor, and Julian Schilliger (ETH Zürich) split the $700,000 Grand Prize after reading over 2,000 Greek letters from PHerc. Paris 4 — a previously unread Philodemus text on pleasure.
February 2025 — Oxford's Bodleian announces the first virtual unrolling of a still-sealed scroll, PHerc. 172. Last ~26 lines of each column recovered.
May 2025 — $60,000 First Title Prize: PHerc. 172 is Περὶ κακιῶν — "On Vices" — by Philodemus. The first title ever read off an unopened ancient scroll.

Project lead Stephen Parsons in 2025: "We're confident we will be able to read pretty much the whole scroll in its entirety." That sentence is the open frontier of 2026.

The library

The Villa of the Papyri was probably owned by Lucius Calpurnius Piso Caesoninus — Julius Caesar's father-in-law. The bias of what we've read so far is overwhelmingly Epicurean, dominated by person philodemus of Gadara (c. 110–40 BCE), who likely lived on the estate.

Only about 30% of the villa has been excavated. The bottom levels — possibly housing a separate Latin collection — were sealed off again in the 1990s for conservation reasons. Hundreds, possibly thousands more scrolls are still down there.

What the rest of the scrolls might contain is the actual prize. Candidates the field talks about, in roughly descending order of plausibility:

More Philodemus, including books of On Vices we know existed by reference
Lost Epicurus — only a fraction of his ~300 reported volumes survives
Stoic, Skeptic, or Academic texts (a few anomalous fragments hint at this)
Early Latin poetry — one fragment looks Lucretian
Long-shot: lost Sappho, Sophocles, Menander, Euripides

The library that survived from antiquity into modern hands came largely through a 9th-century Byzantine copying program. If Herculaneum reads out fully, it would be the first time since then that the classical corpus measurably grows.

What's contested

How much of the surviving carbon-on-carbon signal is real versus hallucinated is the live argument. The Vesuvius Challenge enforces independent verification — multiple teams must produce the same letters on the same patch — precisely because trained models on noisy data can produce plausible-looking nonsense. Some papyrologists remain cautious about specific word-level readings even where the broader text is accepted.

A second contested point: should the rest of the villa be excavated? Italian heritage authorities have resisted further digging since the 1990s on conservation grounds. The opposing camp argues that virtual reading now makes excavation lower-risk — the scrolls no longer have to be unrolled to be read. The decision is political and slow.

A third: even with the full pipeline working, segmentation is still mostly manual or semi-automated. Reading all 1,800 scrolls at current effort-per-scroll is plausibly a decade of work. The bottleneck moved, but it didn't disappear.

Why this has to do with other realms

The ink-detection models are transformer-based vision architectures — the same family of models behind modern LLMs. A neural net designed for the concept transformer architecture problem of predicting text is, with a different head, the thing that's recovering text we lost. The crossover is not metaphorical: several Vesuvius Challenge submissions adapt segmentation models originally built for biomedical imaging, repurposing tumor-detection backbones to find ink.

There's a sharper bridge to concept indus valley script. Herculaneum was a physical decipherment problem — the language was known Greek, the obstacle was substrate. The Indus script is the opposite: the substrate is fine, the language is gone. AI helped where the bottleneck was sensing. Where the bottleneck is meaning, the same techniques have so far produced very little. The two cases bracket what machine learning currently can and can't do for the past.

An open question

If PHerc. 172 reads out in full in 2026, what's the right policy for the unexcavated lower levels of the villa — leave them, or dig? The technology has changed the calculus, but nobody has publicly proposed the new rule.

Key sources

Vesuvius Challenge — scrollprize.org, official progress logs and prize announcements (2023–2025).
Seales et al., 2019 — work on virtual unwrapping of the En-Gedi scroll (PNAS) — the proof-of-concept that preceded Herculaneum.
Bodleian Libraries press release, February 2025 — PHerc. 172 virtual unrolling.
Janko, Philodemus, On Poems (Oxford), for context on what Philodemus's recovered corpus already looks like.
Nature news coverage, 2024–2025, on the Grand Prize and First Title Prize announcements.
To verify: specific Hounsfield-unit figures cited come from Vesuvius Challenge technical writeups — confirm in current scrollprize.org docs before publishing widely.

Abhishek