The AI ‘Maya discovery’ was already published, and the cleanup fell on archaeologists

The Verge reports that Google researcher Pietro Miotti, working on what he called a personal side project, set 48 AI agents to scan NASA lidar data and announced a possibly unmapped Maya village in Mexico. Mexico’s National Institute of Anthropology and History told him it had no registered site at that location. But the site had been documented in an academic journal more than a year earlier, and archaeologists pushed back hard online. Lidar uses lasers to build detailed 3D maps and can reveal structures under forest cover; archaeologists have used it for years.

Miotti said up front he wasn’t an archaeologist, asked the Mexican authorities before going public, said fieldwork would be needed to verify the finding, and updated his post to credit the earlier work once it was pointed out. UCL’s Elizabeth Graham said he did a decent job mapping mounds from lidar. That is more care than many AI discovery claims get.

The failure was the framing. Calling it proof of a revolution across every science after 72 hours skipped the hardest part of research: knowing the literature. As Miotti acknowledged, archaeologists don’t publish site coordinates, so checking a find means comparing images with figures in papers, which his agents couldn’t do. A database with no record isn’t the same as an unknown site. The Verge notes a similar pattern in mathematics, where AI-produced results sometimes duplicate or fail to credit earlier work, part of a wider debate over how AI breakthroughs are checked (our analysis).

Graham says the field’s problem isn’t finding sites; many are known and unexplored for lack of funding, and local communities often already know they’re there. Ivan Šprajc, a co-author of the paper documenting the site, said spotting something adds little to knowledge and that he is already getting more emails from AI “discoverers.” He also flagged a real harm: publicizing site locations can attract looters.

If this scales, AI makes generating claims cheap while checking them stays expensive and falls on experts. Thousands of hobbyists sending agent findings to underfunded researchers means noise, burnout and occasional damage to sites. Security research offers a better model: the labs building AI bug hunters agree the agent that finds a flaw never gets to confirm it (comparison).

The lesson isn’t to keep AI out. Miotti says archaeologists and PhD students have asked about his workflow for their own survey data. That is the useful version: agents helping experts sort huge datasets, with experts setting the questions and checking the answers. More in our research coverage.