Somewhere in the Nazca desert in Peru there are pictures scraped into the ground two thousand years ago — llamas, people, birds — most of them so faint you would walk past them. It took nearly a century of searching to find 430 of these figures. In a survey published in 2024, a team from Yamagata University and IBM Research used AI on aerial photographs of the whole region and, in six months of fieldwork, confirmed 303 more, nearly doubling the count. By 2025 the total found with AI had passed 500, with more than 600 candidates still to check. The non-obvious part is what the machine did and did not do. It produced 47,410 candidate spots. People screened them by eye, processing an average of 36 suggestions to find one promising lead, and spent as many hours at the screen as on the ground. Every figure was confirmed by walking to it, and 59 of the 303 were found by fieldworkers with no help from the AI at all. The AI did not replace the archaeologists; it changed what their days were made of. And the bigger count revealed something the old one could not: two kinds of figures, made for two different purposes.
The Nazca geoglyphs were made between about 500 BC and 500 AD by moving dark surface stones to expose lighter ground beneath. They were spotted by the Peruvian archaeologist Toribio Mejía Xesspe in 1927 and became a UNESCO World Heritage Site in 1994.[3][4] The famous ones — the hummingbird, the monkey — are giant line drawings. Most of the others are small and faint. In the survey paper’s words, “it took nearly a century to discover a total of 430 figurative Nazca geoglyphs.”[1]
The team led by Masato Sakai of Yamagata University, working with IBM Research, trained a model on aerial photographs of the whole region, cut into tiles about 11 metres across. It had only “a few hundred known geoglyphs to train with,” where such models usually need tens of thousands of examples.[1] The model flagged 47,410 candidate spots. Archaeologists examined them by eye and ranked 1,309 as promising. Between September 2022 and February 2023 they walked to 341 locations and confirmed 303 new figures, nearly doubling the known count. The companion page lets a reader try the screening step.[1][6]
The paper is specific about the work this took: “A total of 1,200 labor hours were spent screening the AI-model geoglyph candidate photos. We processed an average of 36 AI-model suggestions to find one promising candidate.” The field survey “required 1,440 labor hours.”[1] Screening and walking took about the same time. The machine made the search possible across the whole plain; it did not make it hands-off.
And it did not find everything. “Of the 303 newly discovered figurative geoglyphs, 178 were individually suggested by the AI and 125 were not.” Of those 125, 66 sat in groups the AI had led the team to; “the remaining 59 were found during the field survey without any help from AI.”[1] Walking to the machine’s suggestions put people in places they would not otherwise have been, and they kept their eyes open.
47,410 AI candidates were screened to 1,309 promising ones; 341 were visited and 303 confirmed on the ground.[1]
The most interesting result is not the number. With nearly twice as many figures, the team could compare the two kinds. The small relief-type figures, about 9 metres across, “depict mainly human motifs or motifs of things modified by humans, such as domesticated animals and decapitated heads (81.6%)” and sit on average 43 metres from ancient walking trails. The giant line-type figures “mainly depict wild animals (64%)” and sit near the ceremonial network of lines.[1]
The paper’s reading: the small figures were “most likely built and viewed at the individual or small-group level,” the giant ones “built and used on a community level for ritual activities.”[1] Two kinds of drawing, two audiences — a distinction that was only a guess with 380 small figures on the map and became an argument with 683. That is the cascade worth noticing: a faster search did not just add entries to a list, it made a question about meaning answerable.
The pace has continued. In November 2025 Sakai told Peru’s state news agency: “With artificial intelligence, we have discovered more than 500 new figures, and we still have over 600 candidates to analyze. It is possible that we could exceed 1,000 figures.”[3] He described the history in three stages — about 30 figures found on foot in Maria Reiche’s era, more than 400 with satellites and drones, and 500 more with AI since 2022.[3] The paper frames the speed as protection: “Climate change poses new threats to Nazca geoglyphs today.”[1] A figure no one knows about cannot be protected from a truck, a flood or a footpath.[4]
Surface stones are moved to reveal lighter ground across the Nazca plain.[4]
The DesertSix months of fieldwork confirm 303 figures from 47,410 AI candidates; 59 were found without AI.[1]
The Survey“It took nearly a century to discover a total of 430 figurative Nazca geoglyphs.”[1]
We processed an average of 36 AI-model suggestions to find one promising candidate. — Sakai et al., PNAS, 2024 [1]
| Dimension | Evidence |
|---|---|
| Operational (D6) Origin · 82 | 47,410 machine candidates → 1,309 promising after human screening → 341 visited → 303 confirmed, in six months of fieldwork.[1]The Search Became a Sieve |
| Quality (D5) L1 · 78 | Every new figure was ground-truthed; 178 were AI-suggested, 66 sat in AI-found groups, and 59 were found without any help from AI.[1]Confirmed on Foot |
| Employee (D2) L1 · 76 | 1,200 labour hours screening and 1,440 in the field: “an average of 36 AI-model suggestions to find one promising candidate.”[1]As Many Hours at the Screen |
| Regulatory (D4) L2 · 64 | A World Heritage Site since 1994, damaged by a truck in 2018; the paper names climate-driven floods as a new threat.[1][4]Can’t Protect What You Can’t See |
| Customer (D1) L2 · 58 | The public account credits the machine; the paper credits a combination of model, screening and walking — and the two kinds of figure it revealed.[1][3]“AI Found Them” Is Half True |
The cascade originates in D6 (Operational): how a survey is run, from searching a desert on foot to screening tens of thousands of machine suggestions first.[1] It propagates to D2 (Employee), where archaeologists spent 1,200 hours at screens and 1,440 on the ground,[1] and to D5 (Quality), where every figure was confirmed in the field and 59 came from people, not the model.[1] It lands in D4 (Regulatory), the protection of a World Heritage landscape under new threats,[1][4] and in D1 (Customer), the public, whose picture of “AI discovering” things omits the screening, the walking and the finds the AI missed. D3 (Revenue) is not scored.
This case begins a thread on AI in archaeology, alongside the library’s cases on AI in mathematics, where the same pattern held: producing candidates got cheap, and confirming them did not. Here the confirmation is physical — a person standing over the figure. The honest limit: the 2025 figure of more than 500 comes from an interview, not a new peer-reviewed count; and the purposes of the two kinds of figure are the authors’ interpretation of where they lie and what they show.
-- UC-335: Thirty-Six to Find One: 6D Amplifying Cascade
-- Nazca geoglyphs, 500 BC-500 AD; spotted 1927 (Mejia Xesspe); UNESCO 1994; Yamagata/IBM 143 new (2019). Sakai et al., PNAS 121(40) e2407652121 (Sep 2024): 430 figurative geoglyphs known after nearly a century; AI on ~11 m tiles trained on a few hundred examples; 47,410 candidates to 1,309 high-potential to 341 visited to 303 confirmed (Sep 2022-Feb 2023); 36 suggestions per promising candidate; 1,200 h screening, 1,440 h field; 178 AI-suggested, 66 in AI groups, 59 without AI; relief-type (~9 m, 81.6% human-related, ~43 m from trails) vs line-type (64% wild animals, near ceremonial lines). Andina (Nov 12 2025): Sakai - more than 500 found with AI, over 600 candidates left.
FORAGE thirty_six_to_find_one
WHERE ai_candidates_screened_by_people = true
AND field_confirmation_required = true
AND discoveries_nearly_doubled = true
ACROSS D6, D2, D5, D4, D1
DEPTH 3
SURFACE thirty_six_to_find_one
DIVE INTO sieve_not_finder
WHEN screening_matches_field_effort = true
AND finds_without_ai >= 59
TRACE survey_work_shift_cascade
EMIT thirty_six_to_find_one_signal
DRIFT thirty_six_to_find_one
METHODOLOGY 90
PERFORMANCE 46
FETCH thirty_six_to_find_one
THRESHOLD 1000
ON MONITOR CHIRP high 'An AI-assisted survey nearly doubled the known Nazca figures in six months: 47,410 machine candidates, 36 screened per promising lead, 1,200 hours at screens and 1,440 on foot, 303 confirmed, 59 found without the AI. The larger count revealed two kinds of figure made for two purposes.'
SURFACE analysis AS json
Runtime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
The model narrowed a desert to candidates; people did the finding. Thirty-six looks for every promising lead, then a walk to each one.[1]
1,200 hours at screens and 1,440 on foot. AI changed what an archaeologist’s day consists of, not whether it is needed.[1]
59 of 303 figures were spotted by people on their way to the machine’s suggestions. The AI put them in new places; their eyes did the rest.[1]
Doubling the count turned a guess into an argument: small figures by the trails for individuals, giant ones for community ritual.[1]
The primary source is the peer-reviewed PNAS paper by Sakai and colleagues, read in full through its open-access copy. The 2025 count comes from the project lead’s interview with Peru’s state news agency; site history comes from reference sources. Source 6 is an interactive companion visual made for this case.
In this case, AI did not remove the work; it moved it — from searching to screening and confirming, in roughly equal measure. The same shift is happening in audits, reviews and investigations. Book a discovery call to map where your organisation’s hours went, and whether anyone is still walking the ground.