How AI is Revolutionizing the Discovery of Ancient Archaeological Sites

Sitting in front of grainy satellite footage of the Peruvian desert, there is a moment when it ceases to resemble geography and begins to resemble a message. Stretching across the Nazca Pampa, lines carved into pale earth are invisible to those who haven’t been instructed where to look and are typically overlooked on maps.

For many years, it took months of fieldwork, skilled vision, and a certain obstinate willingness to crouch in the dust under a harsh sun to find these marks. An algorithm now completes the task in a fraction of the time. Additionally, it’s discovering things that skilled eyes overlooked.

It’s difficult to ignore how subtly this change has occurred. The field of archaeology, which most people still associate with dusty brushwork and khaki-clad professors, has been incorporating artificial intelligence into its fundamental operations in ways that don’t always make headlines but are gradually changing our understanding of the ancient world.

The change isn’t as dramatic as Hollywood had envisioned. It is data-driven, methodical, and, in many respects, more thrilling as a result.

Since 2006, Masato Sakai, the head of a geoglyph research team at Yamagata University in Japan, has been using satellite and aerial imagery to survey the Nazca desert. His team discovered 314 new geoglyphs over the course of about thirteen years of meticulous manual analysis—a truly amazing accomplishment requiring meticulous human attention. AI was then incorporated into the procedure in 2019. They discovered an equivalent number of new figures in a comparable amount of time, but they did so exponentially more quickly.

At first, the AI identified over 40,000 possible locations. After manually reviewing those, researchers reduced the number of geoglyphs to 1,309 that warranted further investigation, carried out field surveys at 341, and eventually verified 303 previously unidentified geoglyphs. Sakai is open about the fact that the system is still not flawless. “AI is not infallible,” he claims. He points out that the limitations are largely dependent on the caliber of the training data.

Nearly every meaningful discussion taking place in the field at the moment revolves around this tension between the speed AI offers and the judgment only humans can provide. GeoPACHA, a research platform located in the Andes, has been working on an even more ambitious project. The project, which was developed by Vanderbilt University’s Institute for Spatial Research, loaded high-resolution satellite imagery of the Andes into a web application.

In its initial phase, which took place between 2020 and 2021, teams manually scoured 180,000 square kilometers of land, documenting what archaeologists refer to as “loci”—individual remnants of ancient human presence. It is nearly impossible to imagine the amount of human labor involved.

The project now uses DeepAndesArch, a foundation AI model that was trained on those previous human-labeled results. To put it simply, the outcomes have been remarkable. Over a million archaeological loci have already been found in less than 5% of the project’s intended total area. The project’s director, Steven Wernke, likens DeepAndes, the model that powers it, to a large language model that was trained on enormous volumes of online text.

However, in this case, the training material is the landscape itself, including its contours, vegetation cover, soil composition, and minute indications of ancient disturbance. “Our foundation model is a latent expert of the land forms and land cover of the Andean world,” Wernke states.

Perhaps the most remarkable feature of this work is the humility ingrained in the methodology rather than the numbers. Wernke takes care to avoid overselling. Some locations are too small or too dim to show up in satellite imagery. Determining the chronological context, or when something was constructed or utilized, is still incredibly challenging from above.

Although the AI can identify the existence of ruins, it is not always able to determine which century they belong to. Archaeologists who walk the ground, touch the stones, and ask the slower questions still own that interpretive layer.

Beyond finding sites, AI is performing tasks in archaeology that, when put simply, seem almost unreal. In order to reconstruct a 12th-century peasant soldier, a team excavating at Miranduolo, a small hilltop site approximately 180 kilometers northwest of Rome, fed ChatGPT excavation data, eliminated anachronistic knowledge, and created an interactive figure named “Johannes” whose face was rebuilt from his real skull. Apparently, you can talk to him. It’s still unclear if this is performance, scholarship, or something completely different.

RePAIR, a 2021 initiative involving the Istituto Italiano di Tecnologia, is reassembling fresco fragments recovered from Pompeii using AI-guided robots. Before a specially designed robotic arm fits the pieces together with the kind of accuracy human hands seldom maintain over hours of delicate work, the software evaluates each broken piece according to size, shape, and decorative pattern.

These initiatives, which are dispersed throughout disciplines and continents, seem to be collectively redefining what preservation even entails.

There has been a similar upheaval in the translation of ancient texts. When a 21-year-old researcher used artificial intelligence (AI) to decipher a portion of a burned and delicate scroll that had not been opened in more than 2,000 years—a scroll thought to be connected to Julius Caesar’s library—he made headlines around the world.

The ancient Greek word for purple, “porphyras,” which he assisted in recovering, was hardly discernible. Depending on where you stand in the field, it is either exciting or unsettling that a young researcher using AI tools achieved what centuries of classical scholarship had not.

It’s hard not to feel that archaeology has reached a turning point that it didn’t quite anticipate as you watch all of this happen. Shovels, notebooks, and patience were all part of the discipline’s gradual evolution, and digital tools were introduced gradually enough to be assimilated. However, AI is advancing more quickly.

Using satellite imagery and LiDAR, the EAMENA project, which covers endangered heritage sites in twenty Middle Eastern and North African countries, is documenting ruins that are under threat from environmental change, development, and conflict. No prior generation of archaeologists had the resources to even attempt the scope of what is being done, which is to record cultural history more quickly than it is being destroyed.

It’s still unclear if AI will eventually take over the interpretive tasks that now call for the intuition of a skilled scholar or if it will always be a potent but subservient tool. Sakai’s advice is measured: remember that the quality of what comes out is only as good as the quality of what went in, approach AI results critically, and refrain from blindly trusting them. That is not cynicism. The genuine accomplishments feel more real because of this kind of honesty.

Silently, the ancient world left its mark in the sand of the desert, under the canopy of forests, and inside scrolls that were too delicate to handle. Finding those marks required being in precisely the right place at precisely the right time for the majority of recorded history. Something is currently simultaneously scanning the entire map.