An artificial-intelligence system trained to recognize Sardinia’s Bronze Age stone towers has produced the kind of result that matters more than a spectacular accuracy score: archaeologists went outside and checked what it had flagged. In a preliminary field campaign across northwestern Sardinia, the team identified six locations they consider probable archaeological sites that do not appear in the regional records.
The result is promising, but it is not six confirmed discoveries carved into the map. The researchers did not excavate the locations, some observations were made from a distance, and five high-scoring spots could not be inspected at all. The real payoff is the documented test: 24 visits, six probable sites, three possible ones, ten negative results, and a wonderfully humbling list of false positives that included a rock-covered landfill and a modern garden wall.
The targets were towers found nowhere else
The monuments are called nuraghes, prehistoric dry-stone towers found only on Sardinia. They vary enormously in preservation and setting. Some remain unmistakable landmarks; others have collapsed into low stone heaps, disappeared beneath vegetation, or become tangled with quarries, towns and newer buildings.
That inconsistency makes them a difficult target for conventional image recognition. A perfect example of a tower is easy to label. A few ancient stones half-hidden by trees can look like almost anything—especially from a satellite image.
The team, led by Augusto Palombini and colleagues at Italy’s CNR Institute of Heritage Science, named its system AIOH, or Artificial Intelligence Operator for Heritage. Their peer-reviewed paper in Archeologia e Calcolatori describes a vision-transformer model designed to examine satellite imagery for the visual and landscape patterns associated with nuraghes.
Five months of labels taught the system what not to see
The training set, called NuragAI, contained 1,612 Google Maps images showing a visible or barely visible nuraghe and 6,751 images without one. Building it took about five months of manual labeling. An archaeologist reviewed the negative examples, and the researchers deliberately included misleading modern shapes such as round buildings and water towers.
On the held-back test images, AIOH reported an average precision of 99.93 percent. That number sounds nearly magical, which is exactly why the paper treats it cautiously. The authors note that the training and test images may be too uniform to represent the messy variety of an entire island. A model can excel on a familiar-looking test set and become much less impressive when the landscape stops cooperating.
The regional experiment was therefore more revealing. The team divided a roughly 60-by-30-kilometer area around Porto Torres, Olmedo and Stintino into grid cells and assigned each cell a probability. Every precisely located known nuraghe in those municipalities lay within 250 meters of a cell scoring above 80 percent, but some monuments themselves landed in lower-scoring cells. Poor preservation and inaccurate historical coordinates may explain part of that gap.
The model may be reading the landscape, not only the stones
The probability maps suggested AIOH was responding to more than a circular ruin. It may also have learned characteristics of places where Nuragic builders tended to put towers. That could be useful: a badly damaged monument might be visually weak while its surrounding terrain still resembles known archaeological settings.
It can also create convincing mistakes. Among the negative field checks were stones covering landfill and a recent dry-stone fence. Those failures are not embarrassing footnotes; they are the evidence readers need to understand what the tool can and cannot do.
At the six probable unrecorded locations, the team saw indicators such as pottery on the ground and marks consistent with worked stone. Four were inspected at closer range, while two were judged from a distance because the researchers lacked permission to enter or dig. Three additional locations were classified only as possible sites. Five could not be checked, and ten produced negative results.
Six leads are not six excavated nuraghes
None of the six probable locations appears in the regional archaeological maps, according to the researchers. They say discussions with local authorities are underway. Until specialists secure access, document the structures in detail and, where appropriate, excavate, the careful word is probable.
That restraint makes the experiment more interesting, not less. This was not a software demo that ended with colored boxes on a screen. The team exposed its model to an unfamiliar landscape, followed its suggestions in person and reported both the archaeological clues and the rubbish piles.
The known tower in the featured photograph is Su Nuraxi at Barumini, Sardinia’s most famous Nuragic complex. It is a contextual example of the kind of monument AIOH was trained to recognize, not one of the six probable new sites. The actual locations have not been publicly identified, an appropriate precaution while their status remains unresolved.
For archaeology, that is the practical promise: not an algorithm that declares history discovered, but one that gives people a better list of places worth checking. The machine points. Archaeologists still have to walk the ground, recognize a garden wall, and decide whether the stones are really telling an older story.
Sources: Palombini et al., “AIOH: Artificial Intelligence Operator for Heritage,” Archeologia e Calcolatori 37.1 (2026); Ancient Herald, November 15, 2026.
Featured image: Su Nuraxi at Barumini, Sardinia, photographed by Norbert Nagel. Wikimedia Commons, CC BY-SA 3.0. Resized to 1280 × 854 pixels; no other changes. This is a known Nuragic complex used as contextual imagery, not one of the probable new sites.