AI Agents Did Not Discover a New Maya Site

Introduction

The ability to locate an unusual pattern is not the same as making a scientific discovery. That distinction became clear after a project involving Google researcher Pietro Miotti used 48 AI agents to search NASA lidar data for possible Maya archaeological sites. One location was presented as a potentially unregistered settlement. Archaeologists soon pointed out that the same site had already been documented in an academic paper.

What happened?

Miotti said the idea emerged while he was traveling in Peru and thinking about how historical explorers might use modern technology. Because publicly available lidar coverage in Peru was limited, he shifted the project to the Maya region. The agents searched for regular shapes that might indicate villages, houses, or other structures. Miotti described the work as a personal side project, not an effort supported by Google computing resources.

After the system flagged a possible large village, Miotti contacted Mexico’s National Institute of Anthropology and History, or INAH. He said the institution had no official record of a registered site at the location. That response, however, could not establish that the site was unknown. Archaeological databases do not include every field observation, unpublished project, or scholarly publication. Coordinates are also often withheld to protect sites, so checking a database is not enough to identify prior work.

Key points

• Detection is not discovery. Lidar and AI can highlight unusual terrain, but experts must determine whether it represents architecture, natural features, or something else.
• Unregistered does not mean unknown. A site may be familiar to local communities, previously visited by researchers, or discussed in scholarship without appearing in an official registry.
• Literature review is part of the discovery process. Archaeologists emphasized that peer review and comparison with existing research are essential for establishing novelty and credit.
• Publicizing locations can create risks. Archaeologist Ivan Šprajc warned that amateur claims may burden specialists and that exposed coordinates could potentially encourage looting.

Why it matters

The episode does not show that AI has no role in archaeology. Archaeologists acknowledged that lidar interpretation can be useful and that automated tools may help specialists prioritize areas for further investigation. The problem is the leap from identifying a pattern to claiming a breakthrough.

A dependable scientific workflow must connect model outputs with publications, maps, unpublished records, expert knowledge, and local context. It also needs field verification before a site can be meaningfully interpreted. As Elizabeth Graham noted, the central challenge is often not finding more sites but securing the funding needed to map, excavate, and understand the ones already known.

The broader lesson applies beyond archaeology. AI systems can produce impressive results while missing the research those results depend on. If amateurs publish unverified discoveries, the burden of checking novelty, correcting attribution, and assessing risks falls on specialists. Future AI research tools should therefore include source tracing, systematic literature checks, expert review, community awareness, and safeguards for sensitive locations. AI may accelerate searching, but discovery still requires interpretation, verification, and responsibility.

Source: The Verge AI