The Herculaneum Scrolls — Reading the Dead with AI
In 79 AD, Mount Vesuvius erupted and buried the Villa of the Papyri at Herculaneum under a pyroclastic surge reaching 300°C. The villa’s library — approximately 1,800 papyrus scrolls, the only intact library from antiquity — was carbonized into fragile black cylinders. They were rediscovered in 1752, and they’ve been illegible ever since.
In 2023, the Vesuvius Challenge launched to change this. By 2025, AI had read the title, author, and passages from a still-rolled scroll for the first time in 2,000 years.
Freshness: 2025–2026 — this is an actively evolving story.
The Buried Library
The Villa of the Papyri was a luxury estate, possibly owned by Lucius Calpurnius Piso Caesoninus — Julius Caesar’s father-in-law. Its library is the only one from the classical world to have survived intact (if illegibly).
What we know about the collection:
- ~1,800 scrolls discovered; hundreds more likely remain underground (only ~30% of the villa has been excavated)
- Overwhelmingly Epicurean philosophical texts — heavily dominated by Philodemus of Gadara (c. 110–40 BCE), a Greek Epicurean poet and philosopher who apparently lived at or near the villa
- Some Latin texts including early poetry fragments
- At least 10 books in Philodemus’s On Vices and their Opposite Virtues series were present
The collection was physically unrolled throughout the 18th–20th centuries — a process that destroyed many scrolls and left thousands of tiny fragments. The still-rolled scrolls were declared too fragile to unroll. They sat in Naples and Oxford, unread.
The Vesuvius Challenge
In March 2023, Dr. Brent Seales (University of Kentucky) and Silicon Valley investors launched the Vesuvius Challenge — a prize competition offering $1.7 million total to teams that could use AI to virtually unroll and read the scrolls from CT scan data alone.
The innovation Seales pioneered: volumetric X-ray CT scanning of the carbonized scrolls at ~4 μm resolution, followed by machine learning models trained to detect the faint ink signal (different X-ray attenuation) embedded in the carbon-on-carbon background. The ink and the papyrus were nearly indistinguishable — but not quite.
Timeline of Breakthroughs
August 2023 — Luke Farritor (21-year-old University of Nebraska computer science student) detected the first Greek letters in a scroll fragment: the word πορφύρας (porphyras, “purple”). He won the $40,000 First Letters Prize.
October 2023 — Youssef Nader (Germany) independently found additional letters, confirming the approach works across multiple scroll fragments.
February 2024 — The $700,000 Grand Prize was awarded to a team of three students — Youssef Nader, Luke Farritor, and Julian Schilliger (Switzerland) — who successfully read over 2,000 Greek letters from a single scroll (PHerc. Paris 4). The scroll revealed philosophical text on the nature of pleasure, consistent with Epicurean philosophy. Key passage: “…as to pleasure, it is clear that…” — possibly discussing whether friendship or material goods provide deeper pleasure.
February 2025 — Oxford’s Bodleian Libraries announced the first virtual unrolling of a still-rolled scroll (PHerc. 172), generating the first images of its interior after ~2,000 years. The papyrological team recovered approximately the last 26 lines of each column — more recoverable text than in any previously scanned Herculaneum scroll.
May 2025 — The $60,000 First Title Prize was awarded for recovering the title of PHerc. 172 noninvasively — the first time a still-rolled scroll’s title has ever been identified without physical unrolling. The papyrological team read:
- Author: Philodemus
- Title: Περὶ κακιῶν — “On Vices”
Deciphered words include ἀδιάληπτος (foolish), διατροπή (disgust), φοβ (fear), βίου (life) — consistent with Philodemus’s characteristic moralistic voice.
Goal as of 2026: Grow from passages to entire scrolls. Project lead Stephen Parsons: “We’re confident we will be able to read pretty much the whole scroll in its entirety.”
What Might Be Lost — and Found
The recovery of On Vices is significant but known. What keeps scholars excited is what else might be in the remaining 1,800 scrolls — the collection may include:
- Unknown works of Epicurus himself — only a small fraction of Epicurus’s prolific output survived the Middle Ages
- Stoic, Skeptic, or Academic philosophy contradicting the villa’s apparent Epicurean focus (some anomalous fragments suggest this)
- Early Latin poetry — one fragment appears to be a Lucretius-era poem
- Possible duplicates of lost Sappho, Sophocles, Euripides
- Historical or scientific texts from the Hellenistic world
If the complete library was as wide-ranging as the Villa’s cultural prestige suggests, its recovery could rival or exceed the 9th-century Byzantine copying program that saved most of surviving classical literature.
The Technical Process
- CT scanning at 4–8 μm resolution (Diamond Light Source UK, European Synchrotron ESRF, and particle accelerator beamlines)
- 3D volume reconstruction from thousands of X-ray images
- Virtual segmentation — tracing the physical surface of the papyrus scroll through the CT volume (a computationally hard geometry problem)
- Ink detection — ML models trained on physically unrolled reference scrolls to detect ink-density signatures
- Character recognition — combining segmentation and ink detection to render readable text
- Papyrological interpretation — human scholars reading and contextualizing the recovered text
The ink-detection breakthrough came from noticing that the carbon ink (a mixture of carbon black, gum arabic, and water) has slightly different X-ray absorption than the carbonized papyrus — a difference of ~1–2 Hounsfield units in the CT scan. The signal is smaller than the noise, requiring sophisticated ensemble ML approaches.
Oxyrhynchus: The Parallel Problem
The Herculaneum victory has renewed attention on Oxyrhynchus, an Egyptian rubbish dump excavated from 1897 to present, which has yielded over 500,000 papyrus fragments — and only ~1% of the remaining buried material has been excavated. These fragments (stored at Oxford’s Sackler Library) include previously unknown lines of Sappho, Sophocles fragments, early Gospel texts, and Menander comedies. AI-assisted reading of existing fragments is underway; a systematic excavation of unexcavated material could be transformative.
Cross-Realm Connections
- concept-voynich-manuscript — both involve ancient documents resisting decipherment; Herculaneum is the success story that shows patient + AI = breakthrough
- concept-indus-valley-script — the Indus script remains undeciphered; Herculaneum shows what AI can do when physical substrate analysis (CT, multispectral) is combined with ML
- event-library-of-alexandria — the Alexandrian library was destroyed; Herculaneum is the only ancient library that survived (if carbonized); together they frame the scale of classical knowledge loss
- concept-transformer-architecture — the ink-detection models are transformer-based vision architectures; the same family of models that reads the scrolls also generates modern LLMs
- concept-archaeoacoustics — both fields recover sensory experience from ancient physical records; CT scans “listen” to carbonized scrolls just as acoustic archaeology “listens” to ancient rooms
- tech-antikythera-mechanism — same era (1st century BCE Hellenistic world); the Antikythera Mechanism and the Herculaneum library represent two kinds of ancient knowledge surprisingly close to us
Key Facts
- ~1,800 scrolls recovered; estimated hundreds more remain underground at the Villa of the Papyri
- $1.7 million in total Vesuvius Challenge prize money awarded (2023–2025)
- Grand Prize winners: Youssef Nader, Luke Farritor, Julian Schilliger (2024)
- First word detected: πορφύρας (“purple”), August 2023, by a 21-year-old undergraduate
- First title identified from a still-rolled scroll: On Vices by Philodemus, May 2025
- Ink signal in CT scan: ~1–2 Hounsfield units difference from substrate — smaller than most scanner noise
- PHerc. 172 (Oxford Bodleian): more recoverable text visible than any previously scanned scroll
- 2026 goal: complete scroll reading of PHerc. 172