AI Archaeology Has a Proof Problem
Key takeaways
- Flagging a possible structure and confirming a Maya archaeological site are separate steps.
- LiDAR can reveal terrain beneath forest cover, but it cannot establish a structure’s age or purpose.
- More AI agents do not guarantee independent verification or fewer errors.
- Public data lets amateurs contribute useful leads, which still need expert assessment and, where necessary, fieldwork.
A pyramid-shaped feature on a terrain map is a lead. Calling it a Maya discovery requires much more evidence. The idea that 48 AI agents could do the job in 72 hours turns on what, exactly, “the job” includes.
A terrain map does not come with a history
LiDAR measures distance by timing how long laser pulses take to return. From those measurements, researchers can build a three-dimensional picture of surfaces.
In an airborne forest survey, some pulses pass through gaps in the foliage and reach the ground. Separating those returns from the rest can reveal terrain hidden beneath the canopy. This is not a way to see underground.
AI can help identify regular slopes, stepped forms, or other features that look worth investigating. But resemblance is only the start. A promising shape might be natural terrain or an artifact introduced while processing the data.
Even a correctly identified human-made structure leaves major questions unanswered. What was it used for? When was it built? What connects it to the Maya?
The same distinction applies to the name on the dataset. Using NASA’s public data would establish a data source, not NASA’s endorsement of an archaeological interpretation. Credibility does not transfer automatically from the institution supplying the measurements to everyone interpreting them.
What does the 72-hour clock include?
Suppose an AI system produces a plausible candidate in three days. That could be useful. It would not necessarily mean an archaeological discovery had been completed in three days.
The clock needs a clear start and finish. Did it include preparing the data? Checking previous surveys and site records? Visiting the location? Or did it stop when the software highlighted an interesting shape?
Those are different measures of progress.
“Not on the map” also needs unpacking. A location missing from one map may already appear in archaeological records. Local residents may know it well. New to a particular dataset is not necessarily new to everyone.
Two questions cut through most discovery claims: What was found, and to whom is it new? Without those answers, the elapsed time tells us little about the significance of the result.
Forty-eight agents can still share one mistake
An AI agent is software that works through a task in multiple steps, potentially reading documents and using tools along the way. Several agents can divide up the work.
One could identify candidate features. Another could compare them with existing site records. A third could look for natural processes that might explain the same shapes. That is a possible workflow, not an established account of how the proposed discovery was made.
The catch is that division of labor does not guarantee independent judgment. Agents working from the same data with similar methods may repeat the same error. Agreement becomes less reassuring when every participant shares the same blind spot.
The more useful question is how the system handles counterevidence. Does it actively test explanations other than “pyramid”? Can another reviewer reproduce the observation? What would make the system reject its favorite candidate?
Agent count describes the setup. It does not establish the quality of the conclusion.
Public data opens the search, not the final verdict
Public datasets let people investigate places they could never afford to survey themselves. AI can help organize those materials and narrow the list of features worth examining.
That gives amateurs a meaningful role. A useful lead identifies the location, names the dataset, and explains why a feature appears human-made. Another person should be able to inspect the same material and see the basis for the claim.
From there, the work moves into comparison with earlier surveys, expert interpretation, and field verification where needed. Establishing that a structure exists is one task. Dating it and connecting it to a particular civilization are others.
AI could make the search faster and open it to more people. But when a discovery is measured in hours, ask what evidence existed when the clock stopped. A shape on a screen can point toward a history; it cannot tell that history by itself.
Comments
Loading comments...