Google researcher's AI-agent Maya 'discovery' turns out to be a known, published site

The claim

Earlier this week, Google researcher Pietro Miotti announced that a team of 48 AI agents had spotted what appeared to be an unregistered Maya site in Mexico, according to The Verge. He described the work as a personal side project that used no Google compute or resources, and framed the result as evidence that a revolution across the sciences was imminent. There was one problem: the site was already known, and had been described in an academic journal more than a year earlier.

Miotti told The Verge the idea came to him while traveling in Peru, where he wondered what historic explorers might have done with modern technology. He settled on lidar, a laser-based mapping technique that produces detailed three-dimensional images of terrain, and pointed the agents at NASA lidar data. Because Peru is poorly covered by public lidar, he redirected the search to the Maya region.

The agents flagged many candidate locations, including a major village that did not appear in existing records. Miotti says he then wrote to Mexico's National Institute of Anthropology and History (INAH) — with the agents translating his messages into Spanish — and was told the institute had no registered site at that location. On his blog, he celebrated receiving an official reply from the Mexican government within 72 hours of the original idea, called for the entire planet to be lidar-mapped so agents could find what remains, and acknowledged that fieldwork would be needed to confirm anything.

The oversight

Specialists noticed quickly. Sarah Parcak, an archaeology professor at the University of Alabama at Birmingham, posted on Bluesky that the episode shows why peer review exists, and noted that a site missing from a government database is not the same thing as an unknown site. Other archaeologists accused the project of spreading misinformation and of eclipsing decades of their own work with digital tools.

"I didn't expect it to get this much attention," Miotti told The Verge. He said that as a non-archaeologist he had not appreciated the difference between a site being unregistered and being unmapped, and that expert knowledge "is needed to verify what the agents find." He also identified a structural weakness in his approach: archaeologists deliberately avoid publishing site coordinates, so confirming novelty means comparing imagery against figures in published papers rather than querying a database — a much harder task for agents. He said he updated his posts to credit the original researchers once the duplication was pointed out.

What specialists say the real bottlenecks are

Ivan Šprajc, a professor at the Research Centre of the Slovenian Academy of Sciences and Arts and a coauthor of the paper that documented the site, told The Verge he was not particularly impressed. Archaeologists have used lidar and other remote sensing tools for a long time, he said, and "just spotting something does not add much to our knowledge." He called Miotti's check with INAH commendable but insufficient, since the institute does not hold all information about known sites. He added that he had already received an email from another self-described AI discoverer and worried about being flooded with such claims, along with a risk that publicizing site locations attracts looters.

Elizabeth Graham, a Mesoamerican archaeology professor at University College London, made a funding argument: researchers in the Maya area already know of many unexcavated and unmapped sites, most found on foot, and the real constraint is money for excavation and mapping rather than detection. She also objected to the word "discover," since local communities are often well aware of sites in their area — while still crediting Miotti with decent work mapping mounds from lidar data, which she noted can be tricky to interpret.

Why it matters

The Verge frames the incident as an instance of a broader AI problem: tools can generate impressive-looking results without engaging the existing research they build on, producing claimed breakthroughs that someone else already made, or findings stripped of the context a specialist would supply. In each case the burden of establishing what was actually found shifts onto experts, work that would normally be part of making a discovery in the first place.

The practical lessons generalize well beyond archaeology. A negative answer from a single registry does not establish that something is new; novelty checks often require comparing against published work in formats agents handle poorly; and verification demands domain knowledge, not just throughput. Notably, Miotti says archaeologists and PhD students have contacted him asking whether the agent workflow could help with their own survey data — suggesting the technique has genuine value when paired with the expertise this run lacked.