Pietro Miotti, a researcher at Google, wrote a blog post this month titled My AI Agents (Re)discovered a Maya Site in 72 Hours. The “(Re)” was added later. In the first version, his agents had found a Classic Maya civic centre on a hilltop in Mexico: a levelled plaza, an oval pyramid “at least 8–12 m tall,” a raised courtyard and what might be a reservoir. Mexico’s National Institute of Anthropology and History (INAH) confirmed in writing that “no registered archaeological site was identified at the location given in your report.”
Then archaeologists noticed. The update at the top of the post, dated 11 November, says the site “had already been mapped in Estrada-Belli et al. 2025. The agents missed it while going through the literature.” The Verge ran the story on 12 November with Sarah Parcak of the University of Alabama at Birmingham saying “This is why we have peer review.”
It’s easy to pile on, and I don’t want to. Miotti says plainly that he isn’t an archaeologist and that the find would need fieldwork. He corrected the post quickly and credited the people who had mapped the site. What interests me is the step where the claim went wrong. The agents found the site without trouble. The failure came at the novelty check, and the evidence that seemed to settle it, a heritage register with no entry for that spot, could hardly have settled anything. Archaeologists have measured how little a register tells you.
Seventy-two hours, thirteen years of data #
The diary is short. On 18 October, travelling in Peru, Miotti wondered what Hiram Bingham would have done with modern tools. That night he had agents turn public lidar strips, “about 1,200 km² of jungle,” into images, and “48 agents went through them one by one like very patient archaeologists, looking for pyramids, platforms, and plazas.” He stresses: “There was no special model or classifier involved: the agents literally looked at the pictures, following a precise guide.” On 19 October he asked them to check “which of these villages haven’t been discovered yet” in the literature and archives, and one came back unregistered. He sent INAH a report on 20 October and had a reply on 21 October.
The lidar itself has a longer history. Miotti’s post labels its lidar view “2013” and sends readers to Cook et al. 2013, the paper describing NASA Goddard’s G-LiHT airborne imager. That paper says G-LiHT had already “been used to collect more than 6,500 km2 of data for NASA-sponsored studies across a broad range of ecoregions in the USA and Mexico.” It was flown for ecosystem science, not archaeology. Archaeologists picked it up anyway. Golden et al. 2016 reported a first archaeological assessment of “extensive transects of lidar recently collected by environmental scientists over southern Mexico using the G-LiHT system.” Archaeologists have kept coming back to environmental lidar from this region ever since.
The two papers nearest the site make this concrete. Auld-Thomas et al. 2024, in Antiquity, analysed an environmental lidar dataset from Campeche, collected for non-archaeological purposes, and found urban centres and dense settlement nobody had recorded. They argue that both are common across the central Maya Lowlands. Estrada-Belli et al. 2025 combined public and privately commissioned lidar, including reprocessed environmental lidar from southern Campeche and southern Quintana Roo, to model settlement and estimate a Classic-period population of 9.5 to 16 million across a 95,000 km² area of the central lowlands. That’s the paper Miotti’s update names.
How big 1,200 km² is #
1,200 km² is not a small search for one person on holiday. But the useful comparison is the area archaeologists had already worked through in the same data. To scale, it looks like this.
Run the agents over a small corner of a region that a 2025 paper has already covered, and most of the hits will be sites somebody has already drawn. Finding one that isn’t would have been the surprise. A careful search would have treated “already published” as the default.
Spotting a pyramid is the easy part #
Lidar under forest is hard to read, and archaeologists have measured how hard. Magnoni et al. 2016 took about 48 km² of high-density lidar from central Yucatán and checked it against areas that had been mapped on foot. Three of the authors went through the imagery separately and counted residential platforms and isolated structures larger than 5 × 5 m. They left out pyramids and monumental buildings, which they describe as “clearly identifiable in the lidar imagery.”
Across the whole table, 101 of 315 structures that had been mapped on the ground could be identified in the lidar, which is 32%, and the readers marked 41 false positives. The authors give three causes: bedrock hummocks one to six metres high “which morphologically mimic house mounds,” dense understory that degrades the bare-earth model, and low foundations lost in “the textural noise of the near-ground vegetation.” They also found that each reader had a habit. Some under-identified and some over-identified, “indicating a personal bias.” Their conclusion is that lidar in that part of the Maya region “should be used as a complement to traditional on-the-ground survey techniques.”
Miotti’s agents had the easy case. A plaza with an 8 to 12 m pyramid on a hilltop is the kind of feature that paper leaves out because anyone can see it. The agents never had to tell a house mound from a bump in the bedrock. The run doesn’t show they could, and Miotti’s post doesn’t claim it does. So it tells us little about how well agents read lidar. It shows they can find monuments where monuments are obvious, which human readers could already do.
What a register leaves out #
INAH’s letter was the strongest evidence in the original post, and it was accurate. INAH didn’t have the site in its register. The mistake was treating that as meaning nobody had recorded the site.
Challis et al. 2008 measured this gap in England. They took 107 km² of Environment Agency lidar over the Dove valley, mapped 915 archaeological features and compared them with the county Historic Environment Records, the UK equivalent of a site register.
772 of the 915 features, 84.4%, were “not previously noted in the Historic Environment Record.” Most were agricultural remains, mainly medieval ridge and furrow, which is hardly news to anyone, so this doesn’t mean lidar finds 772 new sites. It means registers are built from what got reported through official channels, and a lot of what specialists know never reaches them. The same study also found the reverse: records the lidar couldn’t see, such as cropmarks, finds, documentary entries and standing buildings.
The Maya lowlands should make this gap bigger. Estrada-Belli and colleagues model settlement across 95,000 km². I doubt any national register logs every plaza group a regional settlement model picks up. Miotti put it himself in his comments to The Verge: “Archaeologists rightly don’t publish the coordinates of sites, so checking whether a site is already known means comparing images against figures in papers, not searching a database.” That’s the right diagnosis. With coordinates withheld and registers incomplete, the novelty check has to be done against maps and figures in papers, and that’s where the agents failed.
What I’d take from it #
I’d read the 72 hours the other way from how the post presents it. Detecting an obvious monument in public lidar was fast because it was easy. Recognising an already published site in a paper’s figures, when the paper won’t print coordinates, is slow because it’s hard. And the public claim of a find rested on a register lookup that the literature has long known to be weak evidence.
If I were building agents for this kind of work, three things would change.
- I’d treat “not in the register” as no evidence of novelty. The Dove valley numbers give a rough prior: most lidar features aren’t registered, and that tells you nothing about whether anyone has studied them. The novelty check has to run against the regional literature, and for the Maya lowlands that means papers like Auld-Thomas 2024 and Estrada-Belli 2025.
- I’d have the agent check published figures, not only text. With coordinates withheld, it has to compare its lidar views with published site plans and settlement maps. That’s a harder visual-matching task than spotting the pyramid, so it’s where the effort should go.
- I’d report detections with a measured error rate. The Yucatán study gives one benchmark: below the monuments, 32% recall with 41 false positives against 101 hits. An agent pipeline claiming to read lidar should be scored against ground-mapped areas the same way. Monuments the paper calls “clearly identifiable” don’t count as a test.
Miotti ends the post by saying “we should probably lidar the whole planet and let the agents discover the rest.” More lidar would be good. But in the region he tested, the slow step came after the data. It was finding out what archaeologists had already published, and not reading a register’s silence as a discovery.
References
- Miotti, P. (2026). My AI Agents (Re)discovered a Maya Site in 72 Hours . Blog post, last edited 11 November 2026.
- Hart, R. (2026). No, AI agents did not discover a previously unknown Maya site .The Verge , 12 November 2026.
- Estrada-Belli, F., Canuto, M. A., Šprajc, I. and Fernandez-Diaz, J. C. (2025). New regional-scale Classic Maya population density estimates and settlement distribution models through airborne lidar .Journal of Archaeological Science: Reports 66, 105288, October 2025.
- Auld-Thomas, L., Canuto, M. A., Velázquez Morlet, A., Estrada-Belli, F., Chatelain, D., Matadamas, D., Pigott, M. and Fernández Díaz, J. C. (2024). Running out of empty space: environmental lidar and the crowded ancient landscape of Campeche, Mexico .Antiquity 98(401), 1340-1358, October 2024.
- Golden, C., Murtha, T., Cook, B., Shaffer, D. S., Schroder, W., Hermitt, E. J., Alcover Firpi, O. and Scherer, A. K. (2016). Reanalyzing environmental lidar data for archaeology: Mesoamerican applications and implications .Journal of Archaeological Science: Reports 9, 293-308, October 2016.
- Cook, B., Corp, L., Nelson, R., Middleton, E., Morton, D., McCorkel, J., Masek, J., Ranson, K., Ly, V. and Montesano, P. (2013). NASA Goddard’s LiDAR, Hyperspectral and Thermal (G-LiHT) Airborne Imager .Remote Sensing 5(8), 4045-4066, 13 August 2013.
- Challis, K., Kokalj, Z., Kincey, M., Moscrop, D. and Howard, A. J. (2008). Airborne lidar and historic environment records .Antiquity 82(318), 1055-1064, December 2008.
- Magnoni, A., Stanton, T. W., Barth, N., Fernandez-Diaz, J. C., Osorio León, J. F., Pérez Ruíz, F. and Wheeler, J. A. (2016). Detection Thresholds of Archaeological Features in Airborne Lidar Data from Central Yucatán .Advances in Archaeological Practice 4(3), 232-248, August 2016.