How Is AI Used in Archaeology and History?

AI is helping archaeologists and historians in a few specific, checkable ways: reading ink inside ancient scrolls too fragile to open, shortlisting buried settlements from free satellite data, turning photographed medieval manuscripts into searchable text, and drafting transcripts of recorded voices. In every case we have reported, people still do the final reading and judging. Some of these tools are open to anyone today; others are early studies.

The people this matters to are easy to picture: a papyrologist waiting on a scroll nobody has opened since AD 79, a curator who cannot cut a manuscript to learn how it was made, a student with a question about medieval spelling, and families who want an elder’s life story in words they can read. Below are the cases we have reported and checked, a handful of examples rather than the whole field.

Reading what cannot be opened

PHerc. 343 has been a lump of carbon since Mount Vesuvius buried it, and unrolling it would destroy it. Erwin Nieuwlaar, a freelance software engineer from the Netherlands, traced one layer inside an X-ray scan and averaged four machine learning ink-detection models, none ever trained on this scroll. The Vesuvius Challenge’s papyrologists then read letters that most of them take as a phrase meaning “everything that happens.” It is a few letters from one small segment, not yet in a peer-reviewed edition, but the code is open for anyone to reproduce. How the first readable letters came back from a sealed Herculaneum scroll.

Palm-leaf manuscripts from South and Southeast Asia hold their history in the leaf itself. A team led by Laura Gallardo and Patrick Huber at Hamburg University of Technology measured leaves with non-destructive synchrotron X-rays and used machine learning to compare them, in a study in npj Heritage Science. The models reliably told how far a leaf had been processed. Place was harder: India, Indonesia and Thailand were classified correctly more than 80 percent of the time, Sri Lanka and Myanmar far less, and the authors say reliable geographic attribution is not yet possible. What X-rays and machine learning can tell curators about palm-leaf manuscripts.

Finding where people lived

Ancient settlements in central Iraq left low mounds called tells. Elena Chiricallo and colleagues at the Istituto Italiano di Tecnologia and Ca’ Foscari University of Venice trained a deep learning model on free Sentinel-1 radar images and a 30-meter elevation model to find and outline those mounds, published in Remote Sensing. Its scores are modest (an average precision of 0.495), but the authors report archaeological review found it reliable, and it pointed to potential sites not yet mapped. Those are not field-confirmed discoveries, and the authors call the work a baseline. How free radar images could point archaeologists to unmapped settlements.

Turning old pages and old voices into text

A photographed page cannot be searched. A team led by Thibault Clérice at Inria used handwriting recognition to turn more than 23,000 Latin and Old French manuscripts into text, the CoMMA corpus, described in a peer-reviewed paper at LREC 2026. It is already open to students, historians and tool builders. The fine print matters: an average character error rate of 9.7 percent, no human correction, and a later reported total of 32,763 manuscripts that differs from the paper’s figure. What the open CoMMA corpus of medieval manuscripts offers readers now.

In New Zealand, five Cantonese community elders recorded nearly 13 hours of life stories. Sidney Wong of the University of Otago and members of the New Zealand Chinese Association ran the open-source Whisper model over them and got a rough transcript in five hours, against an estimated 508 hours by hand. The catch: it worked well mainly on the English, handled Cantonese and Taishanese poorly, and accuracy was measured on one short segment in a preprint. How an AI first draft could bring Cantonese elders’ stories to families sooner.

Sangmin Lee of Yonsei University with Yejin Jeon and David Ifeoluwa Adelani of McGill University and Mila turned the Endangered Languages Archive into ELARSpeech: 30,202 hours of speech in 675 languages, released for non-commercial use with early speech recognition and synthesis models. Error rates remain high for many languages, one big model for all languages did worse than single-language ones, and the voices were judged by non-native listeners. It is an unreviewed preprint, and the authors consulted the archive rather than each language community. A step toward voice technology for underrepresented languages, built from archives.

How close is this to everyday life?

• Usable now: the CoMMA medieval manuscript corpus is open to anyone, errors and all, and Whisper is already producing first drafts of oral history transcripts, though people must correct the non-English words.
• Getting close: reading sealed Herculaneum scrolls. Letters are coming back, the methods are public, and the Vesuvius Challenge has set a $1 million Grand Prize, with a deadline of June 25, 2027, for fully unrolling and making readable another complete scroll using no more than eight hours of human input.
• Years away: radar prospection still needs field checks of the places it flags, palm-leaf provenance needs larger, more balanced samples, and speech tools for endangered languages need testing by native speakers before they can serve a community.

What ties these stories together is getting voices back: a Greek phrase from a sealed scroll, a scribe’s abbreviations, an elder in Aotearoa explaining how his mother and father said “to eat.” We add new cases as we report and check them on the history beat page.