AI agents found a Maya site. Archaeologists found it first

Pietro Miotti, a Google researcher, was traveling in Peru in October when he started wondering what Hiram Bingham, the explorer credited with “discovering” Machu Picchu in 1911, would have done with modern AI tools. He tried it himself that week. He set 48 AI agents loose on public NASA lidar data over the Maya region, had them scan roughly 1,200 square kilometers of jungle for shapes that looked man-made, and went to bed.

The next morning the agents had flagged something: a hilltop with a leveled plaza, an oval pyramid at least 8 to 12 meters tall, a raised courtyard and what might be a water reservoir. Nothing in Mexico’s official register matched the location. Miotti, who described the project to The Verge as personal and done without any Google compute or resources, had the agents draft a report, translate it into Spanish, and send it to Mexico’s National Institute of Anthropology and History. INAH wrote back within a day: no registered site at that location.

“It took less than 72 hours to go from a random thought while traveling to an official reply from the Mexican government,” Miotti wrote on his blog. “This is by far the best time ever to be alive.” He is not an archaeologist, and said so in the post, but still called the result proof that “we’re on the brink of a revolution across every science” and suggested lidar-mapping the entire planet so agents could find the rest.

The part the agents missed

The site had already been mapped. A paper by Estrada-Belli and coauthors, published in the Journal of Archaeological Science: Reports in 2025, documented the same settlement as part of a wider lidar survey of Classic Maya lowlands. Miotti’s agents never found that paper. On November 11, he added an update to his post crediting the earlier work and changing his framing from “discovered” to “rediscovered.” He told The Verge that comparing a lidar image against the figures buried in academic papers, instead of searching a database, is “much harder for agents, and it’s something these tools need to get better at.”

Archaeologists had already made the same point, less gently. Sarah Parcak, an archaeology professor at the University of Alabama at Birmingham, wrote on Bluesky that the Mexican government not having something in its database doesn’t mean it’s unknown. “This is why we have peer review,” she said. Elizabeth Graham, a Mesoamerican archaeologist at University College London, told The Verge that “our ‘problem’ is not discovering new sites” and that “most of us working in the Maya area know about tons of sites not explored or excavated or mapped, most of these found on foot.” She said local communities are often already aware of the sites outsiders claim to “discover,” and that the real bottleneck for the field is money: there isn’t enough funding to excavate or map sites the way there used to be, so one more spotted mound doesn’t move much.

Ivan Šprajc, a professor at the Slovenian Academy of Sciences and a coauthor of the paper that had already documented the site, was the least impressed. “Just spotting something does not add much to our knowledge,” he told The Verge. He also said checking with INAH alone isn’t enough, because the institute doesn’t have access to all possible information, and that he is now bracing for more of these claims: “Just today I have received an email from another AI ‘discoverer,’ so I am a little bit concerned about being flooded by this kind of message.”

	Miotti’s claim (original post)	What was actually true
Site status	Unmapped, previously undiscovered	Published in Estrada-Belli et al., 2025
Verification method	Checked Mexico’s INAH register	INAH register is not the same as the academic literature
Framing	“Discovered” a new Maya site	Independently found the same site archaeologists had already surveyed

My read

The lidar work itself is real. Graham, one of the archaeologists pushing back hardest, still credited Miotti with doing “a decent job” using AI to map the mounds from lidar data, and called the interpretation of lidar data “tricky.” Šprajc disagreed that it amounted to much: archaeologists have used lidar and other remote sensing tools for years, and “just spotting something does not add much to our knowledge.” Both can be true. The agents found real structures in noisy laser data, and that alone doesn’t tell you much without the fieldwork and literature review that comes after.

What didn’t hold up is the claim sitting on top of it. An AI system checking one government database and coming back clean is not the same as checking the literature, and conflating the two is exactly the kind of overclaim that makes archaeologists distrust AI-assisted work generally. Šprajc’s complaint about being “flooded” by amateur AI discoverers is the real cost here: specialists now have to spend time debunking claims that a day of searching Google Scholar would have avoided. That’s a tax on the field, not a gift to it, whatever the intent.

I’d also push back on “best time ever to be alive” as a verdict on 72 hours of work that turned out to be wrong. The honest version of that story is slower: agents are decent at spotting shapes in noisy data and bad at knowing what’s already been published about them, which is a narrower and much less exciting claim. It’s also the one this incident actually supports.

The pattern looks familiar from other corners of AI. We wrote recently about a startup raising money to make AI agents more reliable, where the company’s own benchmark numbers didn’t agree with each other. Confident output and correct output keep turning out to be different problems, and archaeology just gave us a clean, publicly checkable example of the gap.

Note on sources: I read Miotti’s blog post, including his November 11 update, and The Verge’s reporting in full, both of which carry the on-record quotes used above. I confirmed the title and authors of the Estrada-Belli et al. 2025 paper through its listing, but did not read the paper itself behind its paywall, so its findings here are described only as Miotti and The Verge described them, not independently verified against the paper’s text.

The part the agents missed

My read

Sources