How AI Is Changing Archaeology: From Decoding Dead Languages to Finding Lost Cities
The Algorithm in the Dirt
Archaeology is often imagined as a romantic science — dust-covered researchers with brushes and trowels, painstakingly uncovering fragments of the past one grain of sand at a time. The romance is real, but it obscures a fundamental challenge: the archaeological record is vast, and the tools for analyzing it have been, until recently, limited by human processing speed. A single archaeological site can produce millions of artifacts that take decades to catalog. A landscape survey can cover thousands of square kilometres that no team could walk in a lifetime. A collection of undeciphered texts can sit in museum vaults for centuries, waiting for a breakthrough that human linguists never achieve.
Artificial intelligence is changing this. Not by replacing archaeologists, but by giving them tools that multiply what they can see, read, and understand — opening windows into the past that were previously opaque. The results, over the past five years, have been extraordinary.
Finding Lost Cities from Space
The most dramatic application of AI in archaeology is the detection of archaeological sites from satellite and aerial imagery. Traditional landscape archaeology requires teams walking transects across terrain, a method that’s slow, expensive, and limited to accessible areas. Machine learning models trained on known sites can scan satellite imagery covering thousands of square kilometres in hours, flagging potential archaeological features for human verification.
The results have redefined what’s possible in archaeological survey. In 2023, a team led by Takeshi Inomata at the University of Arizona used lidar (light detection and ranging) data analyzed by machine learning algorithms to discover nearly 500 previously unknown Mesoamerican ceremonial sites in the Mexican states of Tabasco and Veracruz. The sites, identified by their distinctive architectural patterns (rectangular platforms oriented to specific astronomical alignments), were invisible on the ground beneath dense vegetation but emerged clearly in the lidar point clouds when analyzed by the algorithms. The discovery doubled the known number of such sites in the region and reshaped the understanding of Olmec and Maya cultural influence.
Similarly, a collaboration between archaeologists at Durham University and computer scientists used AI analysis of satellite imagery to identify thousands of previously unknown archaeological sites in the Arabian Peninsula, including stone structures (“kites” used for hunting) that span entire landscapes and are only visible from above. The traditional ground-survey approach would have taken decades to cover the same area. The AI approach completed the analysis in months, and the on-the-ground verification is ongoing.
The key innovation isn’t the imagery — satellite data has been available for decades — but the algorithms that can process it at scale. These models learn to recognize the subtle topographic and spectral signatures of archaeological features — slight variations in vegetation, soil composition, and surface texture that are invisible to the untrained human eye. The AI doesn’t replace the archaeologist. It points the archaeologist where to dig.
Decoding Dead Languages
Perhaps the most intellectually thrilling application of AI in archaeology is the decipherment of ancient texts. The challenge of dead languages — those with no living speakers and no bilingual inscriptions (“Rosetta Stones”) — has frustrated linguists for centuries. Linear A, the script used by the Minoan civilization on Crete and still undeciphered after more than a century of effort, is the canonical example. The corpus is small, the language unknown, and every attempt at decipherment has failed.
AI is opening new approaches. In 2023, researchers at MIT’s Computer Science and Artificial Intelligence Laboratory developed a machine learning system that can identify the language family of an undeciphered script by analyzing statistical patterns in character sequences — without knowing what the characters mean. Applied to Linear A, the system suggested a relationship to the Semitic language family, a hypothesis that linguists had proposed but been unable to confirm. The result isn’t a decipherment — the script remains unread — but it narrows the search space dramatically.
More dramatically, a team at the University of Kentucky led by Brent Seales used machine learning and computer vision to read text from the Herculaneum papyri — carbonized scrolls buried by the eruption of Mount Vesuvius in 79 CE that are too fragile to unroll. The “Vesuvius Challenge,” launched in 2023 with $1 million in prize money, attracted teams using AI to virtually unwrap the scrolls and decipher the ink from CT scans. In October 2023, a team of three students (Luke Farritor, Youssef Nader, and Julian Schilliger) successfully read the first word — “πορφύρας” (porphyras, “purple”) — from a scroll that hadn’t been read in 2,000 years. By early 2024, they had decoded entire passages. The technology is still developing, but the promise is staggering: hundreds of carbonized scrolls from Herculaneum, containing works of Greek and Roman philosophy, literature, and science, could become readable for the first time since antiquity.
Reconstructing the Unseen
AI is also transforming the reconstruction of fragmented artifacts. Traditional reconstruction is a three-dimensional jigsaw puzzle that can take years for a single object — fitting together hundreds or thousands of fragments by shape, decoration, and wear patterns. Machine learning models trained on complete objects can now suggest fragment matches that human researchers might miss. A team at the University of Haifa used AI to reconstruct pottery from the ancient city of Tel Dor, matching fragments with 98% accuracy in controlled tests.
The application extends to entire sites. Projects like “RePAIR” (Reconstructing the Past: Artificial Intelligence and Robotics) are developing robotic systems that can physically reassemble archaeological materials using AI-guided manipulation — automating a task that currently consumes thousands of person-hours at museums worldwide.
The Limits of Algorithmic Archaeology
For all its promise, AI in archaeology has limits that are important to acknowledge. The models are only as good as their training data, and archaeological data is fragmentary, biased (rich, dry sites are overrepresented; poor, wet sites are underrepresented), and culturally specific. An AI trained on European castles won’t recognize a great Zimbabwe or a Mesa Verde cliff dwelling without careful retraining. The risk of algorithmic bias — AI systems that “see” only what they’ve been taught to see — is real and could systematically distort archaeological knowledge if not managed carefully.
There’s also a philosophical question: is AI doing archaeology or just doing pattern recognition? Digging a site, interpreting a text, reconstructing an artifact — these are acts of cultural interpretation, not just data processing. The AI can tell you where something is or what pattern it matches, but it can’t tell you what it meant to the people who made it. That act of interpretation — the human leap from data to meaning — remains the archaeologist’s privilege and responsibility. AI is making better archaeologists, not replacing them. The distinction matters.
The Machine Learning Beneath the Surface
The technical approaches that power archaeological AI are worth understanding because they’re applicable far beyond the field. The satellite imagery analysis used to find sites relies on convolutional neural networks trained on labeled data — the same technology used in medical imaging, autonomous vehicles, and facial recognition. The text decipherment work uses sequence-to-sequence models and statistical analysis of character distributions — techniques borrowed from natural language processing and machine translation. The artifact reconstruction uses 3D computer vision and geometric deep learning — the same tools that power augmented reality and robotics.
What makes archaeological AI distinctive isn’t the algorithms but the training data and the application context. Archaeological data is noisy, incomplete, and culturally contingent in ways that clean, abundant, “internet-scale” datasets are not. Building AI systems that work reliably on archaeological data requires collaboration between computer scientists and domain experts that most tech companies don’t fund and most universities don’t incentivize. The results — the lost cities, the deciphered texts, the reconstructed artifacts — are the product of that collaboration. They’re also a reminder that the most interesting applications of AI may not be the ones that generate the most venture capital.
