Pietro Miotti corrected his viral AI-agent archaeology claim on November 11, 2026: the Maya site his agents flagged in Mexico had already been mapped in a 2025 paper. The agents found the location in lidar imagery, then failed to find the relevant prior research.
His updated account now says the agents “rediscovered it rather than discovered it.” That changes the central claim of his November 10 announcement, which described 48 agents reviewing roughly 1,200 square kilometers of lidar data and identifying an apparently unreported site.
The episode separates two tasks that the original story bundled together: recognizing a potential archaeological feature and establishing that it is new. Miotti’s correction preserves the reported image-analysis result while withdrawing the discovery claim.
It also exposes a consequential gap in the workflow. The agents were specifically asked to check the literature and archives. According to Miotti, they missed the paper that would have changed the answer.
What the Correction Changes
Miotti’s original account described a hilltop site with features he and the agents interpreted as a plaza, a pyramid, a raised courtyard and a possible reservoir. After the agents flagged it, he contacted Mexico’s National Institute of Anthropology and History, or INAH, to ask whether it was registered.
INAH’s response, as reproduced in Miotti’s post, said:
“no registered archaeological site was identified at the location given in your report.”
That was a statement about the institute’s register. The later correction says the location nevertheless appeared in Estrada-Belli et al. 2025.
The distinction matters because these checks answer different questions:
| Check | What it can establish | What remains unresolved |
|---|---|---|
| Inspect lidar imagery | A location contains shapes worth investigating | Whether those shapes have been interpreted correctly |
| Search INAH’s register | Whether a site is recorded in that register | Whether researchers have already published it elsewhere |
| Match the location against prior research | Whether it appears in the literature | Its full archaeological interpretation and field status |
An empty registry result cannot carry the weight of a comprehensive literature search. In this case, Miotti’s own update identifies exactly that gap: the site was already mapped, but was not yet in INAH’s register.
My view is that the registry distinction is the most useful detail in the correction. An authoritative institution can give an accurate answer to a narrow question while leaving a broader claim unresolved. The mistake comes when the broader claim inherits the authority of the narrower answer.
What the 48 Agents Actually Did
Miotti’s diary places the work between October 18 and October 21, before the public announcement in November. He says he first discussed archaeological searches with Gemini, ChatGPT, Claude and Grok, then turned to publicly available lidar covering the Maya region.
Lidar measures distances using laser pulses. Some pulses pass through gaps in forest cover and return from the ground, allowing terrain beneath the canopy to be represented separately from vegetation. That makes it useful for finding earthworks and structures that ordinary overhead photographs may hide.
According to Miotti’s account, lidar strips covering about 1,200 square kilometers were converted into images. Forty-eight agents reviewed those images using a written guide, looking for pyramids, platforms and plazas while trying to exclude natural hills, sinkholes and roads. He reports using no specially trained detector or classifier.
The agents later examined the raw point cloud, the collection of individual lidar measurements, to check whether the apparent structures were on the ground rather than artifacts of the canopy. Miotti describes forest roughly 10–17 meters tall above the location.
Those are useful checks, but they address the geometry of the candidate. They cannot tell whether an archaeologist has already published it.
The same caution applies to the reported pyramid height of roughly 8–12 meters and the proposed functions of nearby features. These are interpretations in Miotti’s account, not field-confirmed findings presented in the public material. Miotti explicitly says he is not an archaeologist.
His reported turnaround of less than 72 hours also covers more than image inspection: the diary runs from the initial idea to INAH’s reply, including report preparation and correspondence. It should not be read as a measured processing time for the 48 agents.
The public write-up provides no validated detection rate or controlled comparison against an archaeologist’s review. I’d treat it as a documented workflow account, not an archaeology benchmark. The agent count describes the scale of the reported effort; it does not measure its reliability.
The Overlooked Paper Was Already Linked
Sysnote’s correction report adds a particularly revealing detail: the earlier paper was reportedly linked in the original announcement itself.
Sysnote says it checked a November 10 Internet Archive capture of Miotti’s post and found the 2025 study in its further-reading list. The announcement described that research as work using lidar to estimate population in the Maya lowlands. Sysnote identifies page 14 as the location of the previously mapped site.
That finding deserves careful attribution. Sysnote reports examining the archived announcement; this article is analyzing that published account, not claiming a separate archive investigation. Sysnote also says it could not inspect the paper’s full text through its client. Because Miotti has kept the exact coordinates private, public readers cannot independently reproduce the site-to-paper match from the disclosed materials. The identification rests on his correction.
Even within that limit, the evidence supports a clear lesson: citing a relevant paper and checking a candidate against that paper are separate operations. A bibliography shows that a source was included. It does not demonstrate that its maps, figures or spatial data were examined closely enough to answer a novelty question.
The precise search failure remains undocumented. As Sysnote notes, agent transcripts were not published, so there is no basis for deciding whether the agents skipped the paper, read only part of it or failed to match its mapped features to the candidate.
For spatial research, I’d make that match an explicit verification step. Searching for text about an unknown site is a different task from checking whether a particular coordinate already appears inside a published map.
How the Claim Spread
The Hacker News discussion showed 835 points and 308 comments in the captured page. Its headline presented the result as an unmapped Maya pyramid found in 72 hours.
Those figures document substantial circulation on that forum. They are not independent validation of the discovery, and the commenters do not represent all AI users or archaeologists.
The thread included links to prior research and disagreement over responsibility for the overstatement. Commenter nickalaso focused on the failure to establish whether the site had already been discovered. epistasis criticized media amplification and the attention surrounding the AI framing.
Frequently Asked Questions
4 questions
1Did the AI Agents Discover a New Maya Site?
Miotti now says they rediscovered a site already mapped in Estrada-Belli et al. 2025. His November 11 correction withdraws the original novelty claim.
