Pietro Miotti's 48 AI agents didn't find an unmapped Maya site: it was on page 14 of a paper his own post linked

The Maya site that 48 AI agents "discovered" in NASA lidar data had already been mapped. Pietro Miotti, the Google researcher whose post "My AI Agents Found an Unmapped Maya Site" passed a million views on X and topped Hacker News, corrected it on November 11: the site appears in a 2025 paper by Francisco Estrada-Belli, Marcello A. Canuto, Ivan Šprajc and Juan Carlos Fernandez-Diaz, on page 14. "The agents missed it, I missed it, and it wasn't in INAH's register yet."

The awkward detail is where that paper was hiding. When we pulled the original version of Miotti's post from the Internet Archive (captured November 10 at 20:13 UTC, two hours after the X thread went up), its further-reading list already linked it, DOI 10.1016/j.jasrep.2025.105288, with the description "others used lidar to estimate how many people lived in the Maya lowlands". The prior art wasn't buried in a Mexican archive. It was cited at the bottom of the announcement.

What the agents did

Miotti, whose X bio reads "Research @Google Paradigm of Intelligence" and who stresses he is not an archaeologist, describes a three-day diary from October 18 to 21. The agents found public lidar for the Maya region, turned about 1,200 km² of lidar strips into images, and 48 of them went through those images one by one "like very patient archaeologists", following a written guide to tell pyramids, platforms and plazas from hills, sinkholes and roads. "There was no special model or classifier involved: the agents literally looked at the pictures."

Overnight they flagged many candidate structures. Miotti then asked which of them "haven't been discovered yet", searching "both the literature and the archives", and the agents narrowed it to one hilltop site: a levelled plaza, an oval pyramid of at least 8 to 12 m, a raised courtyard and a possible reservoir. They checked the raw point cloud to make sure the shapes were on the ground and not in the canopy, drafted a report for INAH, Mexico's National Institute of Anthropology and History, and translated it into Spanish. INAH replied on October 21 that "no registered archaeological site was identified at the location given in your report."

That reply is what the "unmapped" claim rested on, and it says less than it seemed to. It is a statement about INAH's register, not about the scientific literature. Miotti hedged at the time ("it's not official yet"), but the headline and the thread said "unmapped", and that is how the story travelled: "finds unmapped Maya pyramid in 72h" on Hacker News, "Google researcher discovers unmapped Maya site" in tech press reposts.

The paper that already had it

"New regional-scale Classic Maya population density estimates and settlement distribution models through airborne lidar scanning" was published in the Journal of Archaeological Science: Reports, volume 66 (October 2025), article 105288, and is open access under CC BY-NC-ND. Its abstract says the authors aggregated public and commercial lidar, "including reprocessed environmental lidar data from Southern Campeche and Southern Quintana Roo, Mexico with enhanced structure visibility", and modelled settlement across a 95,000 km² area of the Central Maya Lowlands. Environmental lidar reused for archaeology is exactly the kind of data Miotti's post describes; it also links NASA's G-LiHT instrument and a 2016 paper on that reuse.

We could not check page 14 ourselves: ScienceDirect didn't serve the full text to our client, and Miotti kept the exact coordinates in a confidential annex. The identification rests on his own correction; the DOI had already been circulating in the Hacker News thread a couple of hours before he posted it on X.

What held up, and what didn't

The detection part arguably came out better than the original story implied. A structure that 48 general-purpose agents picked out of rendered lidar images turned out to be one that professional archaeologists had also mapped from lidar, which makes it a confirmed true positive rather than an unverified anomaly. Miotti says "the rest of the story still holds".

What failed was the step that turned a detection into a discovery: the novelty check. The agents were asked to confirm the site appeared nowhere in the literature, said so, and were wrong about a paper that was open access, on the exact region, and linked in the same blog post. Miotti hasn't published the agent transcripts, so we can't say whether they searched for the paper and misread it, or never looked past its abstract. Our guess, and it is only that, is the second: a site's location in a paper like this lives in a map figure, not in searchable text, and "does this paper's settlement map include a plaza at these coordinates?" is a much harder question than "does any paper mention a new site here?"

That is the useful lesson for anyone pointing agents at science. Spotting candidates at scale is now cheap. Proving that nobody spotted them first still means reading the figures of the papers you already cite.