A search powered by 48 AI agents flagged a Maya site in Mexico as potentially undocumented. But archaeologists soon identified a major gap in the claim: the site had already been documented in an academic journal more than a year earlier. The episode shows that finding a feature in data is only one step in archaeological research; establishing what is known about it takes expertise and careful review.
How the search began
Google researcher Pietro Miotti said the idea came to him while traveling in Peru. He wondered how historic explorers might work with modern technology, and chose lidar, a laser-based mapping method that produces detailed three-dimensional images of the environment.
Miotti put a team of 48 AI agents to work examining NASA lidar data for signs of undocumented sites. He said Peru was not well covered by public lidar data, so he shifted the search to the Maya region. The agents flagged several possible locations, including a village that Miotti thought might not appear in existing records.
He contacted Mexico’s National Institute of Anthropology and History (INAH) to ask whether the location was known, including through unpublished research. Miotti said the institution told him it had no record of a registered site there. He described the project as a personal side project, done without Google compute or resources, and said the exchange took less than 72 hours.
A database check did not settle the question
Miotti acknowledged that the location would need fieldwork to verify. Still, he presented the result as a sign of a wider scientific shift and argued that broader lidar mapping could help AI agents find more sites. Archaeologists responding online said the evidence did not support calling this one a discovery.
Sarah Parcak, an archeology professor at the University of Alabama at Birmingham, stressed the need to compare findings with existing research. She wrote, “This is why we have peer review,” and noted that a site missing from a government database is not necessarily unknown or unmapped.
The distinction matters because institutional records are not a complete inventory of archaeological knowledge. Ivan Šprajc, a professor of archaeology at the Research Centre of the Slovenian Academy of Sciences and Arts in Ljubljana and a coauthor of the paper documenting the site, said INAH’s lack of official knowledge was not enough to establish that archaeologists had never seen or explored the location.
Recognition takes more than spotting a pattern
Miotti said he had not understood that an unregistered site could already have been mapped. He also explained why the agents missed the earlier work: archaeologists do not publish site coordinates, so checking whether a location has been documented can require comparing images with figures in research papers rather than searching a database.
That comparison is harder for agents, Miotti said, and is an area where the tools need to improve. When archaeologists brought the paper to his attention, he updated his post and blog to credit their work. He also agreed that expert knowledge is needed to verify what the agents identify.
For archaeologists, lidar itself was not new. Šprajc said researchers have used lidar and other sophisticated remote sensing tools for quite some time, and argued that “just spotting something does not add much to our knowledge.” Elizabeth Graham, a professor of Mesoamerican archaeology at University College London, credited Miotti with doing a decent job of mapping mounds from lidar data while pointing out that interpreting those images can be difficult.
Finding sites is only part of the work
Graham said many researchers in the Maya area already know of sites that have not been explored, excavated or mapped, including places found on foot. She also cautioned that the word “discover” can erase local knowledge: communities may already be familiar with archaeological sites in their area.
She identified research funding as a larger constraint. Even if a tool surfaces more possible locations, researchers still need resources to investigate and map them. More candidates do not automatically produce more archaeological understanding.
The case points to a broader challenge for AI in science. A system can help sift through large datasets and flag patterns, but a result still has to be checked against earlier research and interpreted in context. Without that work, a new-looking result may repeat someone else’s finding, fail to credit it, or leave experts to establish what the result means. Miotti said archaeologists and PhD students had contacted him to learn about the workflow and consider whether it could help with their own survey data. That kind of use puts AI in a supporting role: helping experts search, while people supply the review and context needed to make a claim credible.