AI at Nazca: From Image Candidates to Field Discoveries

A faint shape in a desert photograph becomes interesting when someone asks whether it could be an ancient figure. An AI model can help direct that attention. The discovery becomes more substantial when researchers examine the location, document what survives and relate it to the surrounding landscape.

The Nazca geoglyph surveys offer an unusually clear view of that chain. They show a productive relationship between image screening and field archaeology, alongside a lesson for anyone writing about AI: keep the candidate, the confirmed finding and the historical interpretation visible as separate stages.

This is a look back at research published in 2023 and 2024, with a later university announcement providing context. The dates matter because a striking old result can easily circulate as if it has just happened.

Begin with the landscape and its makers

UNESCO’s description of the Lines and Geoglyphs of Nasca and Palpa places them in Peru’s arid coastal plain. The landscape contains representations of living creatures and plants, imaginary beings and geometric forms. These are works of ancient human cultures, whose relationships with the land remain central to understanding them.

That context changes the question we ask of the technology. Finding a promising shape is valuable because it can help extend a record of human activity. The image classifier is one contributor to that work; the landscape supplies evidence that a cropped picture alone cannot contain.

Ask what “faster” measured

A study published on 1 June 2023 described a deep-learning pipeline for examining high-resolution aerial photographs of the Nasca Pampa. It reported four geoglyph discoveries verified through on-site surveys.

The authors estimated that their approach could make the image-examination task approximately 21 times faster. Their calculation compared a projected 1,633 hours of unaided visual examination across the pampa with about 78 hours of archaeologists checking model-selected candidates.

That was an estimate for a defined screening task, not a measured 21-fold acceleration of every stage of archaeology. Travel, field documentation and interpretation are different activities. Keeping the measured task attached to the number makes the result more informative.

Read the discovery count as a collaboration

The larger PNAS study, published on 23 September 2024, reported 303 new figurative geoglyphs confirmed during six months of field survey. It concentrated on smaller relief-type figures that are difficult to distinguish in imagery.

Its accounting is revealing: 178 discoveries were individually suggested by AI. Another 66 were found within groups to which AI had directed the researchers. The remaining 59 came from fieldwork or screening outside the AI-suggested boxes. Together, these categories make the reported 303.

The paper’s illustrated examples also include added outlines to help viewers see the figures. Those markings are interpretive aids over the photographs; readers should distinguish them from the underlying image evidence.

A useful reading follows the connections between these stages. A suggestion may bring a team to an area where it notices additional figures. Credit for that contribution can be meaningful even when the model did not individually identify every discovery. The categories help explain how the work progressed.

Keep later announcements in their own frame

A Yamagata University announcement dated 28 July 2025, in Japanese, described further AI-supported fieldwork conducted in 2023 and 2024. It reported 248 newly discovered geoglyphs, including 160 figurative examples, taking the reported figurative total to 893 at that time.

Those categories explain why adding all 248 to an earlier figurative total would be misleading. The announcement also proposed that arrangements along paths could have conveyed stories or messages. That is an interpretation of their spatial organisation, distinct from identifying a figure’s presence.

The 2025 announcement is a university report, separate from the two journal papers discussed here. Its dated total should not be presented as a verified inventory for October 2026.

Detection opens questions about meaning

Consider the difference between finding a shape, classifying its motif and explaining its purpose. Each question asks for a different kind of support. A photograph may help establish the surviving form. A historical interpretation must explain why that form, in that location and relationship to others, supports the proposed account.

For a creator preparing an explainer, this suggests a useful writing discipline. Attach each sentence to the kind of evidence it needs. Describe an observed arrangement as an arrangement, and introduce a proposed purpose as the researchers’ interpretation. Avoid letting a confident caption silently turn one into the other.

The same discipline applies to AI-generated illustrations. A polished reconstruction may make an idea easier to imagine, but it should be labelled as an illustration. It cannot supply missing physical evidence about what an ancient maker actually drew.

Three questions for your next AI discovery story

Before sharing a headline or building a business presentation around research, try these questions. They are a practical reading method, not a new archaeological finding:

• What was produced? Identify whether the reported output is an image candidate, a field-confirmed object, an estimated total or a proposed interpretation.
• What was counted or timed? Keep the category, study period and stage of work beside the number. A screening-time estimate and a completed survey answer different questions.
• What work connects the stages? Look for the checking, documentation and reasoning that turn a model’s suggestion into the result being reported.

MaryChuks.com’s guide to reading a research paper critically offers a broader route through those checks. For this story, the key habit is to read the methods and the count definitions before compressing the work into a sentence.

Recognise the value of a well-directed question

For independent creators and business owners, the transferable idea is about where assistance enters a process. If an AI tool proposes which records, images or enquiries deserve attention, decide what still has to happen before a suggestion becomes an accepted result. Make that transition explicit in the workflow and in any claim about its success.

The Nazca case gives that idea a concrete setting. Careful reporting can recognise a substantial technological contribution while showing the fieldwork and interpretation through which it becomes knowledge. The most useful account lets readers follow that path, including the points where the question remains open.

Sources

• Sakai and colleagues: Accelerating the discovery of new Nasca geoglyphs using deep learning, Journal of Archaeological Science (1 June 2023).
• Sakai and colleagues: AI-accelerated Nazca survey, PNAS (23 September 2024).
• Yamagata University: Further AI-supported Nasca survey results, university announcement in Japanese (28 July 2025).
• UNESCO World Heritage Centre: Lines and Geoglyphs of Nasca and Palpa (reference page checked 25 October 2026).