AI Doubles the Nazca Lines — And Rewrites What They Were For

The giants were never the whole story

For a hundred years, the Nazca Lines were defined by their scale. The hummingbird, the spider, the monkey with its spiralling tail — these are the icons that fill museum gift shops and documentary title sequences. They stretch across the Peruvian desert at lengths of roughly ninety metres, carved so deliberately that some remain legible from the ground.

They are also, it now appears, the exception.

A research team led by Masato Sakai at Yamagata University’s Nazca Research Institute, working with IBM, has spent the last few years training an AI image-recognition model on high-resolution aerial photographs of the Nasca plateau. The result, published in September 2024 in the Proceedings of the National Academy of Sciences (PNAS), is a catalogue revision that nearly doubles the number of known figurative geoglyphs. Three hundred and forty-eight new figures were identified in just six months. One hundred and sixty of them are representational — animals, human heads, livestock — bringing the total count of confirmed figurative Nazca Lines to 893.

But the real story is not the number. It is what the new figures look like, where they sit, and what their existence implies about a tradition we thought we understood.

Size tells a different story

The newly discovered geoglyphs average about nine metres across — one-tenth the size of the famous giants. They are shallow relief-type figures, cut into the desert surface rather than formed by removing the dark surface stones to expose lighter subsoil in long continuous lines. They sit on slopes and alongside trails, not in the open centres of the plateau where the giants sprawl.

This is not a minor aesthetic difference. The placement alone suggests a fundamentally different mode of engagement. The large geoglyphs were designed to be seen from above — or at least, that has been the working assumption since the 1930s, when commercial aircraft first crossed the region and pilots reported seeing the figures from the window.

The smaller ones, tucked into the terrain at human scale, were not. Nine metres is roughly the length of a school bus. You do not need a plane to see them. You need to walk past them. You need to be on the ground.

Sakai himself has suggested that the trail-side figures may have functioned as messages to pedestrians — wayfinding markers, territorial signs, or ritual instructions aimed at people moving through the landscape rather than observing it from it. If that reading holds, it reframes the entire Nasca plateau not as an aerial canvas but as a layered environment in which different visual registers operated at different scales and for different audiences.

The method is the discovery

The Yamagata-IBM collaboration did not simply run a model and call it a day. The workflow was iterative and explicitly archaeological at every stage.

First, the team built a large training dataset from high-resolution aerial photographs of the Nasca plateau and surrounding desert. They fine-tuned a pre-trained image-recognition model on that dataset — essentially teaching it to distinguish geoglyph-like features from natural rock formations, erosion patterns, and modern disturbance.

The AI produced a candidate list of 1,309 locations. Sakai’s team then reviewed those candidates, narrowing them down to the most plausible geoglyph sites. After that, they conducted field surveys and drone photography to verify each location physically. The new figures were not accepted on algorithmic confidence alone.

This hybrid method — machine learning for detection, human expertise for validation, drone verification for confirmation — is what makes the result trustworthy and, more broadly, what makes it replicable. Several previous AI-assisted archaeology projects have stumbled over exactly this gap: models that flag everything and historians who flag nothing. The Nasca work flips the ratio.

Sakai called it a game-changer in his PNAS paper. The smaller, more weathered figures blend into the mottled desert surface in ways that aerial survey simply could not resolve before. A human eye scanning from a plane at altitude would miss them. Even a careful ground survey would struggle — the Nasca research area covers roughly 630 square kilometres, and the team had been doing exactly that since the 1990s, combining satellite imagery, aerial photography, and drone footage, with modest incremental returns year after year.

Why Japan matters in this story

The Nasca plateau is in Peru. The PNAS paper is an American journal. But the lead institution is Japanese, and the collaboration runs through IBM Research in New York. This is not incidental.

Japan has been investing strategically in AI-for-humanities infrastructure for several years — the Ministry of Education, Culture, Sports, Science and Technology has funded multiple cross-disciplinary centres that pair machine-learning labs with archaeology, philology, and art-history departments. Yamagata University’s Nazca Research Institute, founded in 2004, sits at the intersection of that policy push and a site that has resisted easy explanation for a century.

Sakai began his Nazca work in the 1990s. He organized his first research team in 2004. He did not switch to AI until 2018, after a senior colleague pointed him toward the possibility. The eight-year gap between human-scale survey and AI-scale survey is worth noting: it is long enough to suggest that the technology was not ready, and short enough to suggest that the team was deliberate about when to deploy it.

The broader implication is that the AI-archaeology pipeline is becoming institutionalized in ways that are harder to replicate in countries without that kind of funding architecture. Peru has the site. Japan has the computational partnership. The US has the journal. That triad is itself a pattern worth watching.

What the new figures are not

It is useful to say what the PNAS results do not claim. The paper does not assert that the Nazca Lines are now fully decoded. It does not argue that the AI has revealed a complete writing system or a unified cosmological map. The new figures are catalogued; their meanings remain open.

What the paper does show is that the existing catalogue was materially incomplete — not by a small margin, but by a factor approaching two. And that the missing half had a different physical character from the half we already knew. The large, centre-plateau giants were always going to be found by aerial survey. The small, slope-side, weathered figures were not. An algorithm trained to look for them changes the discovery curve.

The method matters more than the count

Doubling a catalogue is impressive. What is more impressive is that the doubled catalogue looks internally coherent rather than noisy. The 160 new figurative geoglyphs display the same stylistic vocabulary — livestock, human heads, zoomorphic forms — as the established corpus. Their estimated dating aligns with the pre-Columbian period of the well-known figures. They are not outliers; they are a missing distribution.

That coherence is the strongest evidence that the AI detection is real and not producing artefacts. If the model had been hallucinating patterns, the new figures would likely have looked stranger — or clustered in ways that correlated with image-quality gradients rather than terrain and trail topology.

Why this matters beyond Peru

The Nasca case is a template for a class of problem that archaeologists across the world are beginning to face: sites that are too large, too subtle, or too degraded for traditional survey to resolve, but in which useful signals are embedded in imagery that machines can now read.

The same approach is being applied — or tested — at sites from the Amazonian geomancy networks to the buried settlement patterns of the Yellow River basin. The variable is not the method; it is the training data. Where a site has a well-maintained photo archive, an AI can be fine-tuned quickly. Where it does not, the model is only as good as the first ground-truthing pass.

Nasca has both. That is why this particular case landed with unusual clarity in a top-tier journal. The PNAS publication is less a surprise about what the AI found than a demonstration that the method works when the data pipeline is sound.

The decade ahead

Sakai’s team published the initial 300-plus figure count in September 2024. By July 2025, Yamagata University reported a further 248 discoveries from the 2023–2024 field season, pushing the figurative total to 893.

The work is ongoing. The next question is not whether more figures exist — they almost certainly do — but what the expanded catalogue tells us about how the Nasca people used space. If the trail-side figures were indeed pedestrian-oriented messages, then the plateau was a multimodal landscape: giants for the sky, smaller figures for the ground, possibly arranged in a sequence that directed movement, ritual, or memory through the desert in ways we have only begun to trace.

A century of scholarship started with planes. It is now, increasingly, starting with processors. The lines themselves have not changed. Our ability to see them has.