Pietro Miotti’s November 10, 2026 announcement credited 48 AI agents with finding an unmapped Maya site in Mexico using NASA lidar data. On November 11, he corrected the central claim: archaeologists had already mapped the site in a 2025 paper. The agents had missed it during their literature search. Miotti’s updated account now describes a rediscovery.
The story had already attracted substantial attention. Its Hacker News discussion reached 1,009 points and hundreds of comments, with participants linking the earlier research. Those figures document the announcement’s reach, not independent validation of the workflow.
The correction leaves two useful findings to examine separately: agents apparently flagged structures that archaeologists had also mapped, but the process failed to establish whether those structures were previously unknown.
How the 48-Agent Workflow Worked
Miotti’s account describes work undertaken from October 18 to 21, before the November announcement. He began by exploring whether public lidar could help identify archaeological remains, eventually focusing on the Maya region because suitable data was available.
The workflow used Google’s Gemini, OpenAI’s ChatGPT, Anthropic’s Claude and Grok. According to Miotti, agents converted lidar strips covering approximately 1,200 square kilometers into images, then 48 agents inspected those images for pyramids, platforms and plazas while trying to exclude natural hills, sinkholes and roads. He says no specialized model or classifier was involved: the agents examined the pictures using a written guide. His post describes the process but is not a comparative evaluation of the four AI products.
Lidar helps here because an ordinary aerial image largely shows the forest canopy. Laser pulses passing through gaps in the vegetation can produce returns from the ground, allowing researchers to reconstruct terrain beneath the trees. Rendered terrain images can then expose regular shapes that are difficult to distinguish in satellite photographs.
One flagged hilltop contained what Miotti described as a leveled plaza, an oval pyramid approximately 8–12 meters tall, a raised courtyard and a hollow that might have held water. The agents also examined slices through the raw point cloud to check whether the shapes appeared in ground returns beneath the vegetation.
That second check matters. A suggestive shape in a rendered image deserves investigation, but checking the underlying measurements helps distinguish a ground feature from a canopy-related artifact. It still leaves archaeological interpretation unresolved: Miotti labels the proposed functions speculative, and his account cites no field survey establishing them.
The agents subsequently helped draft a report for Mexico’s National Institute of Anthropology and History, or INAH, and translate it into Spanish. Miotti says he revised the report before sending it.
I’d treat the candidate-screening stage as the most promising part of this account. Inspecting a large image collection is a concrete task that can be divided among agents. The reported turnaround, however, is elapsed time for this particular workflow, not a measured cost-per-discovery or accuracy benchmark.
What INAH’s Reply Actually Established
INAH replied on October 21. Miotti published an image of the response, which said its records contained no registered archaeological site at the location he supplied. He initially took that as evidence that nobody had mapped it before.
Those are different claims.
An official register answers an administrative question about its records. Scientific literature answers a broader question about what researchers have documented. A location can appear in a research paper without appearing in the register being checked.
Miotti’s November 11 update explicitly acknowledges the failed literature check:
“The agents missed it while going through the literature, so they rediscovered it rather than discovered it”
The update identifies the earlier work as Estrada-Belli and colleagues’ 2025 paper. Sysnote’s reporting identifies it as New regional-scale Classic Maya population density estimates and settlement distribution models through airborne lidar scanning, published in the October 2025 volume of the Journal of Archaeological Science: Reports.
According to that report, the open-access study combined public and commercial lidar, including reprocessed environmental lidar from southern Campeche and southern Quintana Roo, to model settlement across approximately 95,000 square kilometers of the Central Maya Lowlands. That broader study area should not be confused with the 1,200 square kilometers Miotti says his agents screened.
The identification of the same site rests on Miotti’s correction. His public post withholds the exact coordinates, and Sysnote states that it could not inspect the relevant full-text page itself. This account therefore should not be read as a fresh, independent coordinate-level verification.
The Missed Paper Was Already Linked
Sysnote reports a particularly consequential detail: an archived version of Miotti’s original announcement already linked the 2025 paper in its further-reading section, before the correction.
That makes this more instructive than a generic failure to find an obscure publication. The relevant research had reached the announcement’s reference list, yet the novelty claim remained wrong.
In the Hacker News thread, participants linked the paper and pointed readers toward pages 13–14. That documents a specific corrective contribution from the discussion. It does not make every comment reliable or turn the thread into an archaeological review.
The unpublished agent transcripts leave the precise failure unresolved. The agents might have failed to retrieve the paper during their novelty check, reviewed only part of it, or failed to connect a mapped feature with the candidate location. The available record cannot distinguish those possibilities.
A map figure creates a different verification task from a text search. A paper may depict a settlement without naming it in a way that matches a search query. Checking whether a candidate lies inside a mapped area, and whether the arrangement of structures matches, requires spatial comparison.
My view is that this is the most consequential document detail: having a relevant paper in the reference list is not evidence that its maps were checked against the claimed discovery. A bibliography shows that a source was cited. It does not show that the source was examined for the particular fact on which a novelty claim depends.
Detection, Registration and Novelty Need Separate Checks
The published sequence supports a more useful assessment than either celebrating an autonomous discovery or dismissing the entire exercise.
Frequently Asked Questions
3 questions
1Did the AI agents discover a new Maya site?
Miotti corrected that claim on November 11, 2026. He says the site had already been mapped in a 2025 archaeology paper, making the result a rediscovery.
