AI reads a 2,000-year-old scroll charred by Vesuvius, without ever unrolling it
Three students won the 700,000-dollar Vesuvius Challenge grand prize by training machine-learning models to read more than 2,000 letters of philosophy from a Herculaneum scroll carbonized in the 79 CE eruption, without physically opening it.
Artificial IntelligenceGNGV Editorial Team5 min readLast reviewed September 28, 2026
For nearly two thousand years, the carbonized scrolls of Herculaneum, a library buried and burned by the eruption of Mount Vesuvius in 79 CE, were considered all but unreadable; too fragile to unroll, they crumbled at a touch. Now artificial intelligence has begun to recover their words without ever opening them. In a result announced by the Vesuvius Challenge, a global competition, three students used machine learning to read more than 2,000 ancient Greek letters from one such scroll.
The method is a marriage of physics and AI. The intact scroll was imaged with high-resolution CT scanning, producing a 3D map of its tightly wound layers. Competitors then trained machine-learning models to "virtually unroll" the papyrus and, hardest of all, to detect the faint traces of carbon-based ink that are nearly invisible against the carbonized page. The winning team, Youssef Nader, Luke Farritor and Julian Schilliger, shared the 700,000-dollar grand prize for deciphering coherent passages well beyond the contest's threshold.
The text itself is a small wonder. The recovered passages appear to come from a philosophical work discussing pleasure, including the role of music and food in a good life, prompting organizers to joke that the first text uncovered was "a 2,000-year-old blog post about how to enjoy life." With hundreds of Herculaneum scrolls still unread, scholars see the potential to roughly expand the surviving body of classical literature.
The caveats are real but encouraging. Reading scattered passages is not the same as recovering entire scrolls, and the work depends on costly scanning, painstaking annotation and continued refinement of ink-detection models that can still err. The challenge has since shifted toward reading whole scrolls and generalizing across the collection. But an open, prize-driven effort that lets students anywhere train AI to resurrect lost ancient voices is a quietly joyful example of the technology serving human curiosity and shared heritage.
