Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarThe Verge→ original

48 AI Agents 'Discovered' a Maya Settlement, but Scientists Described It a Year Ago

Google researcher Pietro Miotti set 48 AI agents loose on NASA lidar scans and announced an unknown Maya settlement. But scientists had described the site in a journal more than a year earlier. Miotti checked only against the registry of Mexico's INAH institute, and archaeologists don't publish coordinates, so novelty has to be checked against figures in papers. A research agent therefore needs a step that cross-checks publications, not just a search for candidates.

AI-processed from The Verge; edited by Hamidun News
48 AI Agents 'Discovered' a Maya Settlement, but Scientists Described It a Year Ago
Source: The Verge. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceThe Verge ↗
Editor-in-Chief
Key metric #1Lidar is laser 3D scanning of terrain, which can reveal mounds beneath forest canopy
Key metric #2The INAH registry has no record of the site, but that doesn't mean it wasn't studied
Key metric #3Co-author Spraitz has already received an email from another AI "discoverer"

Imagine this: you assemble a team of AI agents, give them open data, and in less than three days you receive an official response from a country's government. It sounds like a perfect demonstration of what the technology can do. That is exactly how Pietro Miotti, a Google researcher, presented the story, until archaeologists pointed out the problem: the "unknown" Maya settlement had already been described in a scientific journal more than a year earlier. For anyone building agents for science, analytics, or any search for "something new," this is a ready-made case study.

What Miotti did and where it went wrong

According to Miotti, this was a personal project that used no Google computing resources. The idea came to him during a trip to Peru: he wondered what historical researchers would do if they had modern technology. He chose lidar, a laser scanning method that builds a detailed three-dimensional model of terrain and can reveal ancient mounds and structures under forest cover.

He launched a team of 48 AI agents and tasked them with reviewing NASA's open lidar data. Public data over Peru is scarce, so he switched to the Maya region in Mexico. The agents flagged many promising locations, including a "large village" that, according to the available records, did not appear in any registry. Miotti wrote to the National Institute of Anthropology and History of Mexico (INAH), using the same agents to translate the letters into Spanish. The institute replied that it had no registered site at that location. On his blog, Miotti called it the best time to be alive and urged that the entire planet be covered with lidar so the agents could "help find the rest."

But the site was already known: it had been documented in a scientific paper published more than a year before. Sarah Parcak, an archaeology professor at the University of Alabama at Birmingham, wrote on Bluesky that this is exactly what peer review exists for. In her view, the absence of a record in a government database does not mean the site is unknown or missing from maps.

Why the agents missed the obvious

Miotti acknowledged that his agents and he himself had made a mistake. Archaeologists rightly do not publish the coordinates of monuments, to prevent looting. As a result, checking whether a find is new cannot be done by querying a database; it means comparing scans against drawings and diagrams in scientific papers. This is far harder for agents, and the tools will have to improve. As a non-mathematical analogy: it's like looking for a duplicate function in someone else's codebase, where the code isn't published anywhere but is described only in words in the documentation.

Ivan Spraitz of the Slovenian Academy of Sciences and Arts, co-author of that paper, noted that contacting INAH was useful but not sufficient: the institute does not have access to all the information. The lack of an official record does not mean archaeologists never noticed or studied the site. He added that scientists have used lidar and other remote sensing methods for years, and simply spotting something in a scan adds little to knowledge.

Elizabeth Graham, a professor of Mesoamerican archaeology at UCL, also challenged the term "discovery" itself. Local residents, she said, usually know very well about the monuments near them, and Maya specialists know of many unsurveyed sites found on foot. The real problem is money: far less is now allocated for excavation and mapping than in the past, so the number of sites "discovered" matters much less. At the same time, Graham acknowledged that Miotti handled the interpretation of the lidar well, which is a difficult task.

What this means for developers

The first lesson: agents generate candidates well but answer the question "has someone already done this?" poorly. If you are building a research agent, include a separate step that checks against publications, plus expert review before any loud claims.

The second lesson: the cost of verification shifts onto specialists. Spraitz said he received an email from another AI "discoverer" on the same day and worries about a flood of similar messages. Just as open repositories suffer from automatically generated pull requests, science could be swamped with claims of "finds."

The third lesson: publishing coordinates carries risk. Spraitz allows that, so far, such finds have not led to looting, so there is no problem yet, but the danger is real.

Miotti himself says he did not expect such attention, updated his post, and credited the authors of the original work. He also notes that archaeologists and graduate students have written to him, interested in his approach to his own scan data. Agents may prove a useful filter over large data sets, but the final word stays with the expert. A similar story is being discussed around recent mathematical results from OpenAI: models produce impressive output but do not always account for prior work. The practical takeaway is simple: before writing "discovered," check the literature and bring in a specialist.

Story

Key Questions Answered

What Miotti did and where it went wrong?

According to Miotti, this was a personal project that used no Google computing resources. The idea came to him during a trip to Peru: he wondered what historical researchers would do if they had modern technology. He chose lidar, a laser scanning method that builds a detailed three-dimensional model of terrain and can reveal ancient mounds and structures under forest cover. He launched a team of 48 AI agents and tasked them with reviewing NASA's open lidar data.

Why the agents missed the obvious?

Miotti acknowledged that his agents and he himself had made a mistake. Archaeologists rightly do not publish the coordinates of monuments, to prevent looting. As a result, checking whether a find is new cannot be done by querying a database; it means comparing scans against drawings and diagrams in scientific papers. This is far harder for agents, and the tools will have to improve.

What this means for developers?

The first lesson: agents generate candidates well but answer the question "has someone already done this?" poorly. If you are building a research agent, include a separate step that checks against publications, plus expert review before any loud claims. The second lesson: the cost of verification shifts onto specialists. Spraitz said he received an email from another AI "discoverer" on the same day and worries about a flood of similar messages.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…