AI Agents in Archaeology: A Maya Site Case Study
See how AI agents searched NASA lidar data for a possible Maya site, how the finding was checked, and what this case reveals about AI in research.

AI Agents Found a Maya Site. Here's What That Really Means
AI agents can now help with research that used to take specialist teams, custom software, and days of manual checking. A recent experiment is a good example: a researcher used AI agents to scan public lidar data over the Maya region of Mexico, flag a hilltop with features that looked possibly archaeological, and send a report to Mexico's National Institute of Anthropology and History (INAH). INAH responded within 72 hours that it had no registered site at that location.
That response matters - but it isn't an archaeological confirmation. The shapes still need expert assessment, and if appropriate, a field visit. The key point is simple: AI can surface a lead, but it can't prove the lead is real.
What's more interesting than the headline is how the workflow worked. The agents didn't "discover" a city by inventing a model. They reviewed lidar images using a human-written guide, compared candidate areas with existing records, and helped draft a report. This article breaks down that process, what lidar can actually show, how to validate agent-assisted research, and what teams in science and business can take from the case.
How did AI agents search lidar data for a possible Maya site?
The experiment started with a straightforward question: could modern tools help find ruins hidden under jungle cover? While traveling in Peru, the researcher thought about how Hiram Bingham might have worked with today's technology - and how lidar might help look past trees.
Lidar works by firing laser pulses and measuring how long they take to return. Some pulses reach the ground through gaps in vegetation. By separating those ground returns from points reflected by leaves and branches, researchers can build a model of the land surface beneath the canopy. That model can reveal elevation changes that ordinary satellite images may miss.
The researcher didn't start in Peru because there was less publicly available lidar there, and commissioning a survey would have raised cost and permission issues. For the Maya region, more public data was available. The agents were given lidar data covering roughly 1,200 square kilometres of jungle, and asked to look for shapes that could be human-made.
The workflow described in the account was intentionally simple. The data was rendered as bare-earth images, then 48 agents reviewed images one at a time. They used a guide to look for features such as pyramids, platforms, and plazas - while trying to separate those from natural hills, sinkholes, and roads. The account also says the process didn't use a bespoke classifier or a special detection model. The agents were effectively examining the images directly.
That detail changes how to interpret the result. This wasn't a fully autonomous archaeological survey. A person chose the region, provided the data, set the task, created or refined the instructions, and reviewed what the agents produced. The agents' job was to spread visual searching across many images and then support follow-up research and report preparation.
In practice, it resembles a research assistant team working in parallel - but only if the brief is tightly defined. The upside is speed and coverage. The downside is that repeated reviews don't automatically create independent evidence. If agents share the same blind spots or follow the same flawed instructions, they can reproduce the same error at scale.
What did the lidar images appear to show?
Among the features the agents flagged was a hilltop at the edge of a cliff. In the bare-earth lidar view, the area appeared to include straight edges and sharp corners under dense forest. The researcher described a levelled plaza, an oval-shaped mound interpreted as a possible pyramid, a raised courtyard, and a flat-bottomed hollow that could be a water reservoir.
The mound's estimated height was around 8 to 12 metres. Nearby platforms were described as potential residential structures. These are interpretations of the terrain, not confirmed descriptions of buildings. Lidar can show where the ground has been shaped. It doesn't, by itself, tell you who shaped it, when it happened, or how the space was used.
The account also says the agents looked at the raw point cloud, not only the rendered image. In the reported profiles, returns from vegetation and ground were separated visually. That matters because a shape that looks architectural in a surface model can be caused or exaggerated by vegetation, processing choices, or gaps in the data. Checking the underlying points helps, but it still isn't a replacement for archaeological review.
The proposed interpretation was a small Maya civic centre from the Classic period, roughly dated to AD 250 to 900. If that's right, the plaza might have served as a gathering place, the mound could have supported a temple, and the courtyard might have been tied to local governance. The hollow might have been an aguada (a reservoir for collecting rainwater during the dry season), or it might be a natural - or human-made - depression with another explanation. The surrounding platforms could have supported houses.
The confidence in each piece isn't the same. The terrain features are visible in the data. Their archaeological character is a hypothesis. Their function and date are hypotheses on top of that. An AI reconstruction can make a hypothesis easier to imagine, but it can't add evidence. It should be treated as illustration, not documentation.
This hierarchy is useful outside archaeology too. When an AI system spots patterns in medical images, satellite data, or financial records, you still want to separate what was directly observed from what's inferred. A good report should label evidence, interpretation, and what remains unknown. Without that separation, a plausible story can slide into something that looks confirmed.
How was the AI-assisted finding checked?
After the agents flagged the hilltop, the researcher asked them to look for existing records and literature. The agents reported they couldn't find a registration for the site. The researcher then contacted INAH with a report (translated into Spanish with AI assistance) and offered to support a field check led by the institute.
INAH replied that it had checked its records and found no archaeological site registered at that location. That's an important administrative update: the location wasn't listed in what the institute checked. It doesn't prove the site is brand new, that no one ever studied it informally, or that the features definitely aren't archaeological. The original account also leaves the identification unconfirmed.
This is the kind of boundary that has to stay clear. There are at least three separate questions here: do the lidar data show unusual ground features, are those features archaeological, and has the relevant authority already recorded them? Lidar can help with the first. Expert review and fieldwork are needed for the second. Archival research and consultation with the responsible institution help with the third.
A responsible validation process for agent-assisted research should include checks like these:
- Inspect the source data. Keep the original point cloud, processing steps, coverage, and known limitations. A rendered image is easier to scan, but it can hide context.
- Test alternative explanations. Ask whether geology, erosion, modern roads, forestry activity, or data artefacts could explain the pattern.
- Review relevant records. Look for published research and consult the appropriate authority. An agent's failure to find a record isn't proof that none exists.
- Make uncertainty explicit. Separate observations from interpretations, and avoid presenting speculative functions or dates as if they were established facts.
- Respect legal and cultural authority. In Mexico, archaeological sites are protected by federal law, and fieldwork requires INAH authorisation. An online detection doesn't grant permission to visit or excavate.
The last point isn't just procedural. Archaeological sites are part of living cultural heritage, and publishing precise locations can increase risks like looting or unwanted attention. The sensible next step isn't to post coordinates and encourage amateurs to dig in. It's to let qualified local institutions decide whether a field visit is warranted and how the information should be handled.
What does this case teach us about AI in research and business?
The strongest lesson here isn't that AI agents replace archaeologists. It's that a capable, general-purpose system can help move a question through several stages: pick a useful dataset, review a large visual collection, organise candidate findings, search records, and prepare a first draft of a report. A human still sets the objective and decides what counts as evidence.
You can see the same pattern in lots of organisations. A team might use agents to screen inspection images, triage customer support cases, compare planning documents, or flag anomalies in sensor data. The operational value isn't just that an agent can answer a question. It's that a bounded task can be repeated across a large volume of material - while people focus on the cases that matter.
But speed only helps if the follow-up process can handle what the system outputs. In business, a detection nobody verifies can create more work than it saves. That means a useful system needs a clear handoff: what gets flagged, what confidence or evidence is attached, who reviews it, what happens when reviewers disagree, and how corrections feed back into the next run.
A practical workflow might look like this:
- Choose a narrow, consequential task. Start where inputs and outputs can be described clearly - like flagging candidate structures in a defined set of lidar tiles.
- Give the agents a clear evidence guide. Specify what to look for, what can be confused with it, and what information to record. Vague instructions tend to produce vague results.
- Keep a traceable record. Save the source images, prompts or instructions, candidate outputs, and human decisions. That makes it easier to repeat the process and spot errors.
- Route uncertain cases to a qualified reviewer. Don't treat the model's confidence or smooth explanation as a substitute for domain expertise.
- Measure the whole workflow. Track useful findings, false positives, time saved, reviewer effort, and the cost of mistakes - not just how many items were processed.
There are limits. Agents can miss features, misread terrain, or describe a shape confidently even when it's not archaeological. Searching literature also isn't the same as a complete archival review; records may be unpublished, inaccessible, or indexed in ways the tools can't interpret. Multiple agents can increase coverage, but they don't guarantee independent verification if they rely on the same data, model, or assumptions.
The broader research context makes the case plausible, not conclusive. Work using NASA's G-LiHT instrument shows how airborne lidar can support environmental research, and studies like Golden and colleagues' 2016 paper describe archaeological uses of lidar collected for other purposes. Research by Auld-Thomas and colleagues in 2024 reported a Maya city identified through lidar analysis. Those examples suggest the general method has scientific precedent - but they don't validate this specific hilltop.
For companies, the takeaway is similar: AI agents tend to be most helpful when they speed up a defined step inside a larger process and leave an auditable trail. The tech still has to fit the quality controls, legal responsibilities, and human expertise the work already requires. If it can't, automation may just deliver unsupported conclusions faster.
Conclusion
This case is a strong example of AI-assisted research, but the usefulness depends on keeping the claim precise. The agents searched a large set of lidar images, highlighted a plausible site, and helped prepare a report. INAH found no registered site at that location in the records it checked. Whether the hilltop is a previously undocumented Maya centre is still an open question.
Even so, it's a meaningful result. The AI helped with a broad visual search and made the administrative follow-up easier for one researcher - while leaving the decisive archaeological work to people with the right expertise and authority. For teams considering similar workflows, the real question isn't whether agents can replace specialists. It's whether they can help specialists find the right evidence sooner, with uncertainty and accountability intact.