AI helps read Herculaneum scrolls without opening them

On June 27, 2026, a research team published a computational-imaging and artificial-intelligence advance that makes it possible to read part of the Herculaneum scrolls’ contents without unrolling them by hand.

Reading without touching

The Vesuvius Challenge brings together machine learning, computer vision and geometry to study these carbonised papyri. Its central difficulty is physical: both the ink and the papyrus contain carbon. In conventional X-ray tomography, the two materials create very little contrast. Writing does not appear like dark ink on white paper.

The process combines several stages. First, the scroll is scanned using high-resolution tomography. Researchers then digitally reconstruct folded, compressed papyrus layers, turn them into three-dimensional surfaces and virtually flatten them. On those surfaces, trained models search for extremely weak signals consistent with ink. It is digital unwrapping: the object stays closed while analysis happens on its imaging data.

On 25 June, the University of Kentucky reported new texts, titles and authors recovered from the scrolls, as well as end-to-end virtual unwrapping and reading of multiple examples. That does not mean every papyrus can now be read or that each scanned area is legible. It means the method has crossed an important threshold: image acquisition, surface reconstruction and reading of writing can be connected in one workflow.

A problem of contrast and geometry

A Scientific Reports paper published in June describes another piece of the puzzle: detecting ink from surface topography. Rather than relying only on density differences, which are tiny in these scrolls, the approach studies microscopic variations associated with writing. The work shows why a general model cannot simply be applied to an image: suitable data must be built, severely deformed layers must be segmented, and the signal must be checked to ensure it really corresponds to strokes.

AI does not automatically “translate” an ancient book. It helps locate and enhance evidence of letters hidden in scans. Specialists in papyrology, ancient Greek and history then identify characters, reconstruct words and judge whether a fragment supports a secure reading. A text may have gaps, poorly reconstructed folds or ambiguous signals; presenting that uncertainty is part of the research, not a failure of the method.

An archive opening again

The Herculaneum scrolls are part of an exceptional library preserved by the eruption of AD 79. Reading them could widen access to works and authors that remained enclosed in carbonised material. But the point is not merely to recover striking phrases. Every readable column needs a chain of evidence: image, geometry, ink detection and philological assessment.

That is the most convincing value of the advance. Artificial intelligence does not replace archaeology or philology; it gives them a way to work with an object that cannot be opened without being lost. Turning inaccessible layers into examinable surfaces lets the past ask questions again, without breaking the material that preserves them.