Advances in artificial intelligence (AI) are usually tied to aspirations for the future, yet they may also transform how we think about the past.
Researchers working at the intersection of archaeology and computer science are using AI to study the peoples and civilizations that came before us. From finding archaeological sites using detection algorithms to deploying computer vision to decipher ancient documents, AI is playing an increasing role in uncovering and telling the story of human culture.
Finding things
Lelia Character is an expert in geospatial machine learning at Sam Houston State University in Texas. Character uses remote sensing technologies, such as Light Detection and Ranging (LiDAR), in combination with machine learning (ML) to help people who want to “find something,” she explained. Her projects include identifying caves in Guatemala, shipwrecks around the U.S. coast, ruined Maya structures, and underwater World War II aircraft wrecks.
Character first used AI in 2018 to help a colleague study climate in the time of the Mayan civilization by looking at stalactites in caves in Guatemala. To find caves, Character used LiDAR and Geographic Information System data in combination with a random forest ML model. However, as LiDAR takes images from above, finding cave entrances—which tend not to face upwards—was challenging. Feature engineering became fundamental to her work. Drawing on her experience of caving, Character identified topographical features such as streams, slopes, and elevation that were “likely to be predictive of whether or not there is a cave entrance presence,” she said.
Since then, Character has continued to develop AI methods for archaeology. Using LiDAR and multi-beam sonar data, she worked with the Underwater Archaeology Branch of the U.S. Navy’s Naval History and Heritage Command to develop a transfer learning method for identifying shipwrecks around the U.S coast. In 2024, working with co-authors from the University of Texas at Austin, the University of Arizona at Tucson, and several institutions in Guatemala, she demonstrated the use of a convolutional neural network object detection model to identify ruined structures of Mayan civilizations, which often are obscured under forest canopies, or hard to access due to remoteness and rugged terrain.
To identify Mayan structures, Character selected features such as hill shade, sky view factor, a value representing the proportion of sky visible from a specific point, and elevation data from the Shuttle Radar Topography Mission, a research effort led by the National Geospatial-Intelligence Agency and NASA that gathered high-resolution topographic data of Earth’s land surfaces. The latter proved the most effective, said Character, because “It makes small-scale features and changes in elevation stand out from the larger ones; if you’re looking for things that are human-constructed, that tends to work well.”
Character’s models output pixel coordinates and image tiles showing locations of predicted sites which she manipulates into geospatial data that can then be investigated on the ground. She is currently working with the U.S. Navy to find submerged planes lost during World War II, with the aim of recovering the remains of missing service members. Her results will be published later this year.
Remote sensing is an active research area globally. Teams at Vanderbilt and Brown universities in the U.S. have used AI-assisted satellite data to carry out archaeological surveys in the central Andean cordillera, while a Portuguese team has shown how LiDAR-generated terrain models and vision transformers can be used for fieldwork validation, and an Italian team has used deep learning to detect archaeological sites within the Mesopotamian flood plains.
Deciphering ancient words
Brent Seales, Alumni Professor of Computer Science at the University of Kentucky, has published work on ‘virtually unwrapping’ fragile artifacts using X-ray computed tomography and computer vision. The geometric-based technique involves segmenting, flattening, and texturing three-dimensional scans to produce two-dimensional images that can reveal text—as ink marks—embedded in scrolls. Since 2019, Seales has been applying the method to damaged scrolls from Herculaneum in the Bay of Naples, rolled papyrus documents carbonized during the eruption of Mount Vesuvius in AD 79 and discovered in the library of a villa in the 1750s.
Seales has released scans of thousands of the Herculaneum scrolls as part of the Vesuvius Challenge, a contest launched in 2023 calling on AI researchers to solve technical problems associated with reading the scrolls. It also seeks to devise efficient ways of scanning the hundreds of scrolls yet to be digitized.
Seales explained that ink in the scrolls is particularly hard to decipher, as evidence of the ink exists, but is not visible to the naked eye. “AI became the tool that we used to enhance the evidence of the ink so that it could become visible to the human eye from this method of imaging,” Seales said.
In 2023, Luke Farritor became the first to identify an entire word on a scroll. Farritor, then a 21-year-old computer science student, trained an ML algorithm to search the scroll scans for ‘crackle,’ a surface texture that fellow Vesuvius Challenge contestant Casey Handmer had previously identified as looking like ink. Farritor’s model was refined with each discovery of more crackle, allowing him to eventually identify ink strokes, then letters, and finally the word Πορφυρας, meaning “purple.” A slew of other AI-driven solutions have followed. The Vesuvius Challenge’s 2025 Grand Prize stands at $200,000, to be awarded to the first team that can virtually unwrap an entire scroll and reveal “readable letters” within.
With hundreds more scrolls yet to be scanned, as many as a million words could be revealed, said Seales. “That means that it has to change our understanding of the ancient world.”
Diversity of data
Text poses its own challenges, but archaeology is based on a multitude of different kinds of data found in architecture and objects left behind by humans. According to Peter J. Cobb, a specialist in digital humanities at the University of Hong Kong and digital reviews editor of Advances in Archaeological Practice, ML and archaeology are a good fit, as “analyzing large data sets and looking for patterns is basically what archaeologists do.” Yet the diversity of data is tricky. Said Cobb, “Can we get our data into a form where it would be suitable for a machine learning task? That’s a huge challenge.”
Data can also be incomplete. When teams excavate, they may only be digging 1% or 2% of a site, Cobb explained. “Then you’re only getting what was left behind, what survived in the ground for thousands of years.” Further, data from early excavations is often stored in notebooks rather than digital form, and “There’s no incentive to digitize old journals from 1920s excavations, 1880s excavations,” he said. Older visual data can be limited, too, as the relative cost of film prior to digitization restricted the number of photographs archeologists took. Overall, less data was captured in the past, and it is filtered through the eyes of earlier scholars, said Cobb. “Lots of interpretation had to happen just deciding which data to collect.”
While not all archaeological data is yet fit for training models, the arrival of Generative AI raises new possibles, according to an assessment by Cobb. Like other fact-based disciplines, archaeology is likely to face issues with bias and hallucinations in GenAI. However, Cobb believes it could be useful for automating processes and freeing up time for fieldwork. He also sees applications in outreach, museums, and education, for example, “To create videos that introduce the public to concepts; they can be edited and controlled by archaeologists so that the narrative is correct,” he said.
While AI-powered archaeology is still in what Cobb describes as an “experimental phase,” it has already demonstrated the potential to speed up discoveries and unlock information held within ancient artifacts and sites. While widespread focus on how AI is shaping the future, archeologists will keep finding ways of using it to study the past.
Karen Emslie is a location-independent freelance journalist and essayist.