48 AI agents rediscovered a Maya site already mapped by archaeologists

Forty-eight AI agents flagged a Maya site in laser-survey imagery, but the novelty check failed. Pietro Miotti corrected his personal account on November 11, 2026: archaeologists had already mapped the location in a 2025 paper. The result was a rediscovery, not a previously unknown site.

In Miotti's October experiment, agents examined NASA data and literature. He revised their report and submitted it to Mexico's National Institute of Anthropology and History, mistaking an absence from its register for an absence from previous maps.

Those are two separate questions. Does an image contain a human-built feature? And has somebody studied it before? Spotting a house in an aerial photograph does not establish that you are its first discoverer. An empty entry in one register cannot answer for the entire research literature.

Lidar itself is an established surveying technique. Laser pulses measure distances, producing a three-dimensional account of the landscape. Processing can separate vegetation returns from ground returns. Imagine lifting a green blanket off a scale model: regular platforms and steps become much easier to distinguish from the covering above them.

Archaeologists have been doing this work for years. Tulane University described a 2024 study that identified more than 6,500 pre-Hispanic structures in Campeche, Mexico, including a previously unknown large city. A separate 2025 study combined public and privately commissioned surveys to examine settlements, plazas and agricultural infrastructure across a much wider region. Recognizing a striking mound is only part of that broader interpretive task.

The correction does not show that agents are incapable of reading terrain. It exposes the gap between detecting something interesting and establishing that it is new. Linking image inspection, literature searches and report writing into one automated workflow does not make its weakest stage dependable.

That distinction travels beyond archaeology. People using AI to prepare research notes or reports need the trail of earlier work, not simply a polished conclusion. Extra agents can divide the burden of scanning pictures. A claim that nobody has done something before still needs a substantive comparison with existing evidence, rather than a search that happened to return nothing.

Experiment: October 2026. Author's correction: November 11, 2026.