"My AI agents found an unmapped Maya site," one of the most-discussed posts on Hacker News this week, has been corrected by its author. In a November 11 update to his write-up, Pietro Miotti says the hilltop ruin his agents flagged in NASA lidar data had already been mapped in a 2025 study by Estrada-Belli and colleagues. The agents missed that paper during their literature check. The site was simply absent from the register of Mexico's National Institute of Anthropology and History (INAH), which is why the institute told him it had no record of it.
The correction hasn't caught up with all the coverage. TechRadar's November 11 piece, for example, still describes a site not yet mentioned in the scientific literature. What remains is a smaller but still useful story. One person with no archaeology background, using off-the-shelf AI agents and public data, went from idea to a credible lead sent to a national heritage agency in about three days. The lead was real, just not new.
What he actually did
Miotti, an AI researcher (TechRadar describes him as a Google researcher; his write-up names no employer), lays out the timeline on his site:
| Date | Step |
|---|---|
| Oct 18 | While travelling in Peru, he asks several chatbots how modern tools could find lost sites. They suggest lidar. He picks the Maya lowlands for their public lidar coverage and sets agents loose that night |
| Oct 19 | Agents report many settlement-like shapes, then check literature and archives. One site appears unregistered |
| Oct 20 | Report refined, translated into Spanish by the agents and sent to INAH, with coordinates in a confidential annex |
| Oct 21 | INAH replies that its records show no registered archaeological site at that spot |
| Nov 10 | He posts the story publicly |
| Nov 11 | Update: the site was already in Estrada-Belli et al. (2025), making it a rediscovery |
The method was simple, which is the interesting part. Roughly 1,200 square kilometres of NASA lidar strips over jungle (10 to 17 metres of canopy, per CellCog's reading) were converted into images. Then 48 agents looked at the pictures for pyramids, platforms and plazas, following a written guide that told them to rule out natural hills, sinkholes and roads. Miotti says there was no special model or trained classifier: the agents "literally looked at the pictures." Candidates were checked against the literature and archives, then against the raw point cloud to confirm the features were ground, not vegetation.
What they flagged, by his description: a cliff-edge hilltop with a levelled plaza, an oval mound at least 8 to 12 metres tall, a raised courtyard, a flat-bottomed hollow that might be a reservoir, and three platforms that are probably house foundations. He speculates it was a Classic-period (roughly AD 250–900) civic centre. He labels that interpretation as speculation that came out of a long chat with the agents and allows that the hollow may be nothing more than a hole.
What isn't known
- Which model ran the agents. The write-up doesn't say, and doesn't give compute cost.
- How many of the other flagged shapes were already known. Only the one site was pursued.
- Ground truth. No one has visited. Fieldwork in Mexico needs INAH authorization, and Miotti says the decision rests with the local INAH centre. He has offered to help.
- Outside expert review. None of the coverage we read quotes an archaeologist assessing the find. CellCog, which sells AI research agents and discloses that conflict, notes the same gap.
Why the correction is the lesson
Lidar has already turned up thousands of Maya structures. In 2024, Tulane's Luke Auld-Thomas and Marcello Canuto reported in Antiquity that lidar collected for other purposes revealed more than 6,500 pre-Hispanic structures over about 50 square miles of Campeche, including a previously unknown city. Miotti himself points to that study. With that much already published, the bottleneck isn't spotting mounds in elevation maps. It's knowing what the specialist literature has already recorded. That is exactly the step where his agents failed, while sounding confident about it.
Who should care: anyone using agents for research. Visual triage at scale worked here. The "is this new?" step didn't, and it's the step a newsroom, investor or grant reviewer will care about most. Treat an agent's "no prior record found" as a search result, not a finding, and have a domain expert check it before you announce anything. Miotti's own suggestion, to lidar-map the whole planet for agents to analyse, will need that expert in the loop too.
Sources
- Pietro Miotti — My AI agents found an unmapped Maya site pietromiotti.com
- TechRadar — Google researcher discovers unmapped Maya site in Mexico using 48 AI agents techradar.com
- CellCog — AI agents find unmapped Maya site: what is confirmed cellcog.ai
- Tulane University Stone Center — Have we found all the major Maya cities? Not even close, new research suggests stonecenter.tulane.edu