ML ink detection reads 2,000+ characters of a Herculaneum scroll

AI system

System: Vesuvius Challenge grand-prize ink-detection models (TimeSformer-based ensemble) applied to virtually unwrapped CT segments
Model: The submission combined a TimeSformer-based model with other architectures.
Release: Exact release not established.
Application developer: Youssef Nader, Luke Farritor and Julian Schilliger (contestants)
Role: The models detected traces of ink in CT images of the scrolls.
Configuration: The model was part of the 2023 grand-prize submission to the Vesuvius Challenge.
Inputs and outputs: Not assessed
Autonomy: Not assessed
Workflow: High-resolution synchrotron CT scan of a carbonized scroll. Virtual unwrapping into flattened segments. Ink-detection model trained on labeled crackle/ink texture. Independent papyrological review of the output.
Human oversight: Not assessed
Relation to AI: The contestants' ink-detection models produced the ink images from which papyrologists read the text. The ink is not visible in the raw CT data. Virtual unwrapping (non-AI) was a necessary precondition. The readings were validated by an independent papyrological review organized by the contest.
AI contribution: The contestants' ink-detection models produced the ink images from which papyrologists read the text. The ink is not visible in the raw CT data. Virtual unwrapping (non-AI) was a necessary precondition. The readings were validated by an independent papyrological review organized by the contest.

Studies: Not assessed
Study design: Not assessed
Measurement: Not assessed
Setting and sample: Not assessed
Measures: recovered 15 columns and more than 2,000 characters of Greek text
Comparison: the scroll as it stood before the 2023 contest, with none of its text read
Results: Reference: 0% of a scroll. Result: 5% of a scroll; 15 columns of never-before-seen text, more than 2000 characters total.
Additional value from AI: Not assessed
Limitations: Not assessed
Denominators: Not assessed
Materials for replication: Not assessed

Kind of value: Vesuvius Challenge grand prize awarded for the winning submission that read the scroll
Amount: $700,000
Currency and prices: USD, identified from the source context. Nominal source amount; price year not separately stated.
Period and unit: Grand-prize announcement on February 5, 2024. One prize shared by a three-person team.
Realization: Awarded.
Recipients and payers: Recipient: Youssef Nader, Luke Farritor, and Julian Schilliger. Payer: Vesuvius Challenge prize fund.
Gross or net: Gross prize award to recipients, not new aggregate wealth or research profit.
Included costs: Not assessed
Comparison: No cash counterfactual is stated.
Method: Not assessed
Return and payback: Not assessed

Participants: Developer: Youssef Nader, Luke Farritor and Julian Schilliger (contestants). Operator: Vesuvius Challenge (co-founded by Brent Seales, University of Kentucky).
Jurisdictions: Operator: United States.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Conditions for replication: Ink must produce a detectable texture or density signal in the CT data. Segmentation must be accurate enough to preserve letter order. Readings must be confirmed by papyrologists, not only by model output.
Replication attempts: 2 follow-up entries. Benefit recurred: 2.
Later findings: 5 February 2025. The same AI ink-detection approach was applied to a second scroll, PHerc. 172 at the Bodleian, scanned at Diamond Light Source in July 2024, yielding columns of text including the word διατροπή.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Assessed benefits: People. Organizations. Society.
Depth: People: 2. Organizations: 2. Society: 3.
Reasons: People: Nader, Farritor and Schilliger shared a $700,000 prize announced by the organizers on 5 February 2024. Farritor had earlier won the First Letters prize for the first word. Organizations: The University of Kentucky program behind the virtual unwrapping, described by Seales as a 20-year dream, was validated with a readable text. Society: A previously unknown Greek philosophical text became readable. Nature reports that the method paves the way to deciphering the rest of the scrolls.
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: Independently reported

Kind: Not assessed
Related entries: Not assessed
Field: Humanities / papyrology. Open competition / academic research.
Purpose: Built to detect ink in X-ray CT scans of carbonized Herculaneum papyri so that the text can be read. This benefit was that purpose.
Dates: Date basis unassessed: 2024 (precision unassessed). Updated: 21 October 2026.
Stage: Not assessed
Version: SHA-256: 69be89d05e9a…