A widely shared experiment using 48 AI agents to inspect lidar imagery did not discover an unknown Maya site, as its creator initially believed. The software engineer behind the project updated his account on November 11 after learning that the location had already been mapped in research published in 2025. The episode still demonstrated how general-purpose AI tools can help screen large remote-sensing datasets, while underlining the importance of a thorough literature review before making a discovery claim.
The engineer said he obtained public lidar covering roughly 1,200 square kilometres of forest in Mexico. Lidar measures reflected laser pulses and can reveal the shape of the ground beneath vegetation. He converted strips of the data into images, then assigned 48 agents to inspect them for regular features associated with archaeological sites, including platforms, plazas and pyramid-like forms. The workflow did not use a purpose-built archaeological classifier; the agents viewed images under a common set of instructions.
One hilltop area appeared especially promising. The imagery showed straight-edged forms around a level space, an oval mound estimated at eight to 12 metres high, a raised courtyard and a depression that might be a reservoir. Those interpretations were preliminary and based on remote sensing, not a field survey. The agents also checked the underlying point cloud to reduce the risk that vegetation had produced the apparent shapes.
The engineer submitted a report to Mexico’s National Institute of Anthropology and History, or INAH. According to his account, the agency replied that it did not have a registered archaeological site at the coordinates. That response encouraged the initial claim that the location was unmapped, but it did not establish that the site was unknown to researchers or verify the proposed interpretation of its features.
A subsequent check found the site in a 2025 publication by Estrada-Belli and co-authors. The engineer accordingly changed his description from a discovery to a rediscovery and acknowledged that the agents had missed relevant literature. The distinction matters: an absence from an institutional register is not the same as an absence from scholarship, and neither remote imagery nor an AI-produced report replaces archaeological assessment on the ground.
The project also illustrates a potentially useful division of labour. Automated agents can divide a large image set, flag unusual terrain and help assemble a report quickly. Human experts and authoritative records remain necessary to assess earlier work, interpret ambiguous formations and decide whether a field visit is justified. Any archaeological work in Mexico requires INAH authorisation, the engineer noted, and no field confirmation was reported in his update.
