Skip to content
Week 46News, and what it means for you

Science

AI agents found a Maya site, but it was already in a study

Pietro Miotti, who is not an archaeologist, had AI agents search NASA lidar images and reported an unknown Maya site in Mexico. On 11 November he added that it had already appeared in a 2025 study.

Pietro Miotti, who says he is not an archaeologist, had AI agents search NASA lidar images and reported a Maya site that was not in Mexico’s register. On 11 November he added that archaeologists had already mapped the spot in 2025.

It began on 18 October in Peru, where Miotti was reading about Machu Picchu. He put the idea to Gemini, ChatGPT, Claude and Grok, and they all suggested lidar: height measurements taken with a laser from an aircraft. He chose the Maya region, because more public lidar exists for it. Some 1,200 square kilometres of survey strips were turned into images, and 48 agents went through them one by one. They followed a written guide: look for pyramids, platforms and plazas, and skip natural hills, sinkholes and roads. Miotti then asked whether any of the settlements were unregistered, and one agent flagged a single site.

It is a hilltop on the edge of a cliff in Mexico, with a levelled plaza, an oval pyramid at least 8 to 12 metres tall and a flat-bottomed hollow that may once have been a reservoir, under 10 to 17 metres of forest. On 20 October Miotti sent a report in Spanish to INAH, Mexico’s national institute of anthropology and history. A day later INAH replied that no registered archaeological site had been identified at that location.

Around 10 November Miotti put the story online. It scored more than a thousand points on Hacker News and was picked up by the tech site RuntimeWire. On 11 November the correction followed: the site already appears in a 2025 study by Francisco Estrada-Belli and colleagues, which the agents had missed while searching the literature. So it was a rediscovery. The site was not yet in INAH’s register, and according to Miotti the rest of the story still holds. On Hacker News one reader called the method “downright primitive” next to how archaeologists work today; others praised it as a good use of AI.

“There was no special model or classifier involved: the agents literally looked at the pictures, following a precise guide.”

Pietro Miotti on his blog

How lidar sees through the forest

Lidar measures height with laser pulses fired from an aircraft. On the way down a pulse bounces off leaves, branches and shrubs, and sometimes part of it reaches the ground. Dense tropical forest blocks up to 96% of the pulses, LIDAR Magazine wrote, but fire enough pulses per second and some still find the gaps. Software then strips out whatever bounced off vegetation, leaving the bare ground with every mound, wall and terrace. At Copán in Honduras it was already clear in 2000 that this makes ruins under rainforest visible. Whatever shows up in the images still has to be checked on the ground.

A revolution in Maya archaeology

In 2018 the PACUNAM project published what it had found in some 2,100 square kilometres of forest in northern Guatemala: more than 60,000 structures, including houses, roads, fields and fortifications. “All of us saw things we had walked over,” said the archaeologist Marcello Canuto of Tulane University.

In 2024 the doctoral student Luke Auld-Thomas found an unknown city in Campeche, Valeriana, in lidar collected in 2013 for a forest survey measuring carbon. He had deliberately looked for such existing data, because new flights are expensive. In July 2025 Estrada-Belli, Canuto and colleagues combined public and private surveys, including reprocessed environmental lidar from NASA, covering some 95,000 square kilometres of lowland. Their estimate: 9.5 to 16 million people between AD 600 and 900. According to Miotti, his site is in that study too.

Known is not the same as registered

INAH’s answer was about the register: there was no registered site at that location. That is not the same as unknown, as this case showed: the site was already in a journal, but not yet in the register. A reader with fieldwork experience pointed out on Hacker News that many site records exist only on paper and are hard to tie to a location.

Our analysis

Agents make the searching fast, but the checking is still human work

What worked is impressive. Agents with a good guide did in hours what would otherwise be days of screening, on public data and without a specialised model. Researchers such as Auld-Thomas and Estrada-Belli work the same way: looking again at existing surveys. The guide made the difference, because it set down what counts and what can be discarded.

Where it went wrong is more instructive. The agents checked whether the site was registered anywhere or appeared in the literature, found nothing, and the report treated it as a new site. But finding nothing in what you searched does not prove it is not there. The site was in a study of the same region, based partly on the same kind of NASA data. An agent reports finding nothing just as confidently as it reports a find. Our expectation: the same mistake will turn up in businesses, with an agent that searches contracts, patents or client files and says there is nothing comparable, while the answer sits in an archive it could not reach.

Miotti added the correction openly within a day. For a director the lesson is simple: let agents do the searching, but treat a conclusion that something is new or missing as a hypothesis, and have an expert test it against all the sources.

What stands out

  • Unregistered is not unknown INAH only replied that there was no registered site at that location. The site had already appeared in a study in a journal.
  • The study used the same kind of data Estrada-Belli and colleagues worked partly with reprocessed NASA lidar, just like Miotti’s agents.
  • Less than 72 hours From a passing thought while travelling to an official reply from the Mexican government took under three days, according to Miotti.
  • The location stayed secret Miotti put the coordinates in a confidential annex. A site on a public map attracts looters, readers on Hacker News pointed out.

What it means for you

Agents can do days of searching in hours, but when they find nothing, that only tells you about what they could read.

  • Write a guide Miotti’s agents worked from a precise description of what counts and what does not. A guide like that makes searching repeatable and checkable.
  • Ask what was searched Have an agent list the sources it read. Then you can see what is missing, such as a paper archive or another system.
  • Have an expert test the conclusion When an agent says something is new, unique or missing, have it checked by someone who knows the field.

Where we come in · Solutions

We deploy AI that shows where the answer comes from

Our AI knowledge base answers questions from your own manuals, instructions and quotes, with the source alongside, so a colleague can always check where an answer came from. And our custom AI agents take over routine work, such as labelling emails or pulling invoice details out of PDFs. We agree in advance what a colleague reviews first.

What searching may AI do for you, and who checks it?

Tell us what your colleagues now spend hours looking for. We’ll think it through with you and always reply within one working day.

More from week 46

All news

Search our services, the news and the rest of the site.