AI Reads the Burnt Vesuvius Scrolls: Decision Support, Not Magic

For nearly 2,000 years, hundreds of papyrus scrolls carbonized by the eruption of Mount Vesuvius in 79 AD have been unreadable — too fragile to unroll without destroying them. Since 2023, the Vesuvius Challenge has changed that, not by cracking the scrolls open but by training machine-learning models to find ink inside high-resolution X-ray scans. It is a textbook case of AI as decision support: the models reveal where letters likely are, and human papyrologists, classicists, and the competition’s review team decide what the text actually says. AI surfaces the signal; scholars own the meaning.

Share this reportXLinkedInFacebookReddit

Key takeaways

  • A 2023 prize with humanist roots: The Vesuvius Challenge launched in March 2023, founded by Nat Friedman, Daniel Gross, and University of Kentucky computer scientist Brent Seales, building on two decades of “virtual unwrapping” research.
  • First word read by a student: In October 2023, 21-year-old Luke Farritor became the first person to read a whole word from inside an unopened scroll — ΠΟΡΦΥΡΑϹ (“purple”) — winning the $40,000 First Letters Prize.
  • A grand prize for real passages: In February 2024, Youssef Nader, Luke Farritor, and Julian Schilliger shared the $700,000 Grand Prize for recovering more than 2,000 characters across roughly 15 columns of Greek text.
  • A title, finally: In 2025, scroll PHerc. 172 was identified as On Vices by the Epicurean philosopher Philodemus — the first time a sealed Herculaneum scroll’s title and author have been read.
  • Humans stay in the loop: Every claimed reading is independently reviewed by the challenge’s papyrological team before it counts — AI proposes, scholars verify.
~800
still-rolled Herculaneum scrolls awaiting reading
Source: Herculaneum papyri / Vesuvius Challenge, 2024
2,000+
characters recovered for the 2024 Grand Prize
Source: Vesuvius Challenge, 2024
$700K
2024 Grand Prize shared by three winners
Source: NEH, 2024
~$1.8M
total prize money awarded to date
Source: Vesuvius Challenge, 2025
79 AD
year Vesuvius carbonized the Villa of the Papyri library
Source: Herculaneum papyri, historical record

The problem: a library no one could open

The Herculaneum papyri are the only large library to survive from classical antiquity — more than 1,800 scrolls recovered in the 18th century from the Villa of the Papyri, with roughly 800 still rolled and carbonized into brittle lumps of ash. Physical unrolling destroyed many in earlier centuries. The deeper obstacle is the ink itself: it is carbon-based, sitting on carbonized papyrus, so it has almost no density contrast in an X-ray scan. You cannot simply “see through” the scroll and read it. That is precisely the gap where machine learning earns its place — detecting patterns far too subtle for the human eye, then handing them to experts. Source: Herculaneum papyri (Wikipedia) and ink-detection research, 2023–2024. Reference

The method: scan, unwrap, detect, verify

The pipeline has four stages, and only one is fully automated. First, intact scrolls are imaged with high-resolution micro-CT at particle-accelerator facilities such as the Diamond Light Source near Oxford. Second, software “virtually unwraps” the scroll — segmentation tools like Brent Seales’s Volume Cartographer and Julian Schilliger’s ThaumatoAnakalyptor trace each rolled layer and flatten it into a readable surface. Third, ink-detection neural networks (trained on the faint “crackle” texture that carbon ink leaves in the papyrus) highlight where letters probably sit. Fourth — and decisively — human papyrologists read, transcribe, and cross-check the output. Brent Seales’s team spent two decades proving each step was even possible, including virtually unwrapping the En-Gedi scroll in 2015. Source: The Conversation / University of Kentucky, 2024. Read more

2023–2024: from one word to whole passages

In October 2023, Luke Farritor — then a 21-year-old computer science student and SpaceX intern — trained a model that surfaced the first readable word from inside a sealed scroll: ΠΟΡΦΥΡΑϹ, Greek for “purple.” It won the $40,000 First Letters Prize, which required finding at least 10 legible letters in a 4 cm² area. Four months later, in February 2024, Farritor joined Youssef Nader and Julian Schilliger to win the $700,000 Grand Prize, recovering more than 2,000 characters across roughly 15 columns of philosophical Greek text. The winning entry combined automated segmentation, ink detection, and expert interpretation — and every reading was vetted by the competition’s papyrological review team before the prize was awarded. Source: National Endowment for the Humanities, 2024. Read more

2025: a sealed scroll names its own author

In May 2025, the challenge announced its most striking literary milestone yet: scroll PHerc. 172, held at Oxford’s Bodleian Libraries and scanned at the Diamond Light Source, was identified as On Vices by the Epicurean philosopher Philodemus — possibly the first book of his ethical treatise. It was the first time a still-rolled Herculaneum scroll’s title and author had ever been read. The find was made by Würzburg graduate students Marcel Roth and Micha Nowak (with challenge researcher Sean Johnson independently reaching the same reading), and it was confirmed by the papyrological team before Roth and Nowak received the $60,000 First Title Prize. Again the shape holds: software made the letters visible, classicists determined what the letters meant. Source: CNN / Vesuvius Challenge, 2025. Read more

What’s next: scale, not just spectacle

The frontier now is throughput. Reading one scroll was a breakthrough; reading roughly 800 of them is an industrial problem of faster scanning, more robust automated segmentation, and cheaper ink-detection models — work the challenge continues to fund through open prizes. The open-source, open-data design matters here: more than 1,000 teams have competed, and shared code and scans compound progress. But scale does not change the division of labor. Even if every scroll is virtually unwrapped, the result is a flood of candidate Greek text that only trained papyrologists and classicists can authenticate, edit, and place in intellectual history. Source: Vesuvius Challenge, 2025. Read more

The through-line: decision support, not replacement

At every milestone — the first word in 2023, the first passages in 2024, the first title in 2025 — AI did the same narrow job: it revealed where ink probably is on a surface no human could otherwise read. It never decided what the text says, what it means, or whether the reading is trustworthy. Those judgments stayed with papyrologists and the challenge’s independent review board, who must confirm each result before any prize is awarded. The Herculaneum scrolls are a clean demonstration that the highest-value AI in scholarship is not an oracle but an instrument — a tool that expands what experts can examine while leaving interpretation, verification, and authority firmly in human hands.

Methodology & sources

  • Challenge launched March 2023 by Friedman, Gross, and Seales; over 1,000 teams; ~$1.8M total prizes — Vesuvius Challenge (2024)
  • Farritor read the first word ΠΟΡΦΥΡΑϹ (“purple”) and won the $40,000 First Letters Prize, October 2023 — Scientific American (2023)
  • 2024 Grand Prize of $700,000 to Nader, Farritor, and Schilliger for 2,000+ characters across ~15 columns — NEH (2024)
  • PHerc. 172 identified as Philodemus’ On Vices; first title read; $60,000 First Title Prize — CNN (2025)
  • ~800 still-rolled scrolls of 1,800+ from the Villa of the Papyri; carbon-based ink — Herculaneum papyri (2024)
  • Virtual unwrapping, segmentation, and ink-detection methodology — The Conversation (2024)

Frequently asked questions

Does the AI actually read the scrolls, or do humans?

The AI detects ink; humans read it. Machine-learning models highlight where carbon-based letters likely appear in X-ray scans of virtually unwrapped scrolls, but papyrologists and classicists transcribe and interpret the text, and the Vesuvius Challenge’s independent review team must verify every reading before a prize is awarded.

How can a scroll be read without being opened?

It is scanned with high-resolution X-ray CT and “virtually unwrapped” in software. Particle-accelerator facilities like the Diamond Light Source image the intact scroll in 3D; segmentation tools trace each rolled layer and flatten it digitally, and ink-detection models reveal the faint texture left by the ancient ink — all without physically unrolling the fragile papyrus.

What has actually been read so far?

The first word (“purple”) in 2023, more than 2,000 characters of Greek philosophy in 2024, and a scroll’s title and author in 2025. That last milestone identified scroll PHerc. 172 as On Vices by the Epicurean philosopher Philodemus — the first time a sealed Herculaneum scroll’s title has ever been recovered.

Part of our Real-World AI Use Cases series — how AI supports high-stakes decisions across surprising domains.