An AI Nazca discovery helped archaeologists confirm 303 previously unknown figurative geoglyphs in Peru’s desert during a six-month field survey. The result nearly doubled the number of known figures in the study area after almost a century of conventional searching. Artificial intelligence did not authenticate the drawings by itself. It analysed aerial imagery, ranked promising locations and helped researchers decide where to look. Archaeologists then examined the ground and drone images before adding any figure to the inventory.
The distinction is essential because the word discovery can hide several stages. A model can flag a shape that resembles a geoglyph, but wind, paths, geology and modern activity can produce misleading patterns. The published work combined computer vision with repeated human inspection. That partnership allowed a small research team to search an enormous landscape quickly without treating every machine prediction as an archaeological fact.
How the AI Nazca Discovery Worked
Researchers from Yamagata University, IBM Research, the German Aerospace Center and other institutions began with high-resolution aerial images of the Nazca Pampa. Their system searched for small relief-type figures, which are made by moving stones to expose lighter ground beneath. These are harder to identify from the air than the enormous line-type figures most people associate with Nazca.
The model produced 47,410 candidate boxes across the survey region. Researchers visually reviewed those suggestions and narrowed them to 1,309 higher-potential candidates. Field teams then visited 341 locations during work conducted in 2022 and 2023. According to the PNAS study, they confirmed 303 new relief-type figurative geoglyphs during that six-month survey.
Not every confirmed figure maps neatly to one box selected by the model. Some were direct AI candidates, some belonged to groups identified around a candidate, and others were noticed by archaeologists while they were already surveying a model-prioritised area. The accurate claim is therefore that AI accelerated and guided the search. It did not independently see, classify and verify 303 drawings from a screen.
Yamagata University’s official project account says the AI-assisted approach increased the discovery rate sixteenfold compared with earlier fieldwork. That figure concerns the efficiency of finding candidates worth checking, not a claim that the model was sixteen times more accurate than an archaeologist. Human experts remained the final authority at every confirmed site.
Why 303 New Figures Could Remain Hidden for So Long
The Nazca landscape contains geoglyphs on radically different scales. Giant line-type designs can extend across hundreds of metres and include animals, plants and geometric forms. Small relief-type figures may be only a few metres across, have faint outlines and sit close to paths. Erosion, disturbed stones and overlapping traces can make them difficult to distinguish in a vast aerial image.
Before this survey, researchers had documented 430 figurative geoglyphs over almost a century. The new work concentrated on the smaller category that conventional visual scanning was least likely to find efficiently. This is a classic computer-vision problem: the useful signal is sparse, variable and surrounded by thousands of natural or human-made shapes that look similar at first glance.
LiveAIWire’s earlier overview of AI archaeology and digital dig sites described how machine learning can search remote-sensing data at a scale no field team could match. The Nazca project shows the mature version of that workflow. The model did not replace excavation or interpretation. It compressed a huge search space into a manageable set of places where expert time was most likely to produce evidence.
The New Drawings Were Mostly About People
The new inventory changed more than the total. Among newly catalogued relief-type motifs, 81.6 per cent depicted humans or subjects closely connected with human activity. They included humanoid figures, heads, domesticated camelids and scenes involving people and animals. Decapitated heads also appeared, imagery already known in the region’s material culture.
That pattern differs from the giant line-type figures. In the researchers’ broader comparison, 64 per cent of those large designs depicted wild animals, including birds, felines and other creatures. The contrast suggests that the two forms were not simply large and small versions of the same artistic practice. Their subjects, placement and likely audiences appear to have differed.
The paper also reports a striking relationship with movement through the landscape. Relief-type figures sat, on average, about 43 metres from winding trails. Giant line-type figures were associated instead with formal networks of straight lines and trapezoids, with an average separation of roughly 34 metres. These distances were calculated from mapped features, giving the interpretation a spatial basis rather than relying only on what the pictures resemble.
Some Figures May Have Worked Like Messages Beside a Path
The team argues that small relief-type figures were likely seen by individuals or small groups walking along informal trails. Their human-centred subjects may have communicated information about activities, identity or ritual to people moving through the area. Yamagata University’s summary compares their function to signs or display boards, a useful analogy but not a literal description left by the people who made them.
The giant line-type geoglyphs appear to fit a different setting. Their connection with straight lines and trapezoidal networks is consistent with community-level ceremonial routes. The researchers infer that groups may have viewed or used them during organised ritual activity. That remains an archaeological interpretation based on form and distribution, not direct testimony about what every design meant.
These conclusions are valuable precisely because the AI widened the sample. A handful of newly found figures might add curiosities. Hundreds can reveal statistical differences between categories. The system’s main scientific contribution was therefore not recognising one spectacular animal. It helped create a larger, more systematic inventory from which archaeologists could test ideas about how the landscape was used.
The Desert Preserved an Ancient Visual Landscape
UNESCO says the Lines and Geoglyphs of Nasca and Palpa cover about 450 square kilometres of Peru’s arid coastal plain, roughly 400 kilometres south of Lima. The designs were made between about 500 BC and AD 500 by removing dark surface material or arranging stones so that lighter ground created a visible contrast. The dry environment helped preserve marks that would have disappeared quickly in wetter terrain.
Preservation does not make the site static. Vehicles, foot traffic, development and erosion can damage shallow surface features. A catalogue can therefore support conservation as well as research. Once a geoglyph is mapped and verified, authorities have a better chance of monitoring change, limiting accidental disturbance and considering it in land-use decisions.
That protective role also creates a dilemma. Publishing precise locations can help scholarship while increasing exposure to damage or looting. LiveAIWire’s reporting on AI archaeology finding sites experts missed examines the same tension: faster detection must be paired with local stewardship and careful decisions about what location data should become public.
Why Human Verification Still Decides What Counts
Computer vision is good at ranking resemblance, not establishing cultural origin. A candidate may have the right outline but lack the stone displacement, surface weathering or relationship with other features that would support identification as an ancient geoglyph. Field inspection can examine those details from several angles and compare them with regional archaeological knowledge.
The Nazca team used drones and ground surveys to document candidates before confirmation. That procedure resembles the verification chain used in other forms of scientific imaging. LiveAIWire’s account of how AI helped read a closed Herculaneum scroll likewise separates machine assistance from expert interpretation. In both projects, AI exposes a possible signal. Specialists decide whether the underlying physical evidence supports the claim.
The model also inherits the limits of its training examples. It is more likely to recognise forms resembling geoglyphs already known to researchers. Unusual construction methods, heavily damaged figures or motifs absent from the training data may remain invisible. A ranked-candidate system improves search efficiency, but it does not prove the areas it ranks low contain nothing important.
The Discovery Is a Map for Future Archaeology
The immediate next step is to examine the remaining high-potential candidates and continue comparing the distribution of relief and line types. Better imagery and improved models may reveal fainter traces, but each new prediction will still require evidence on the ground. The most responsible measure of progress is not how many boxes an algorithm produces. It is how many defensible discoveries survive archaeological review.
The work also points towards broader uses of AI in cultural heritage. Models can help prioritise landscapes for survey, compare old and new imagery for damage, and organise collections too large for manual inspection. LiveAIWire has explored how AI is digitising cultural heritage, where the same principle holds: scale becomes useful only when the system preserves provenance and keeps qualified people in the decision loop.
The 303 figures matter because they transform a famous archaeological site without pretending the machine solved its mysteries. AI supplied speed and focus. Drones supplied detailed images. Archaeologists supplied confirmation and interpretation. Together, those layers turned faint marks beside ancient paths into a much richer picture of how people represented animals, ritual and themselves in one of the world’s most extraordinary designed landscapes.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.