How AI Is Helping Archaeologists Reconstruct the Past

The past rarely arrives intact. More often, it comes as a scatter of sherds, a handful of eroded letters, or a faint crop mark in a field. For centuries, archaeologists have stared at such fragments and tried, through immense patience, to reconstruct what once was. Today, they have a new collaborator. Artificial intelligence is not digging trenches or dusting off bones, but it is doing something increasingly important. It is learning to detect patterns and see the invisible connections that human eyes can easily miss. This is not a story about replacing the archaeologist. It is about amplifying the archaeologist’s ability to restore a shattered world, piece by digital piece.

Reading Broken Texts Nobody Could Finish

One of the most dramatic applications is in ancient writing. The Vindolanda tablets — thin wooden writing tablets found near Hadrian’s Wall — include hundreds of fragments so faded that scholars have spent years trying to read them properly. Infrared photography, multispectral imaging, digital microscopy, and computer-assisted interpretation have helped make some of these difficult texts more legible. They have revealed the everyday world of Roman Britain: military orders, private letters, requests for supplies, and even messages about socks, sandals, and underclothes. Small details. Huge windows.

The Herculaneum scrolls are an even more striking case. Buried under volcanic material from the eruption of Vesuvius in AD 79, these papyri are so fragile that attempts to unroll them physically can destroy them. Researchers and competitors in the Vesuvius Challenge have used X-ray tomography, computer vision, and machine-learning methods to detect ink within still-rolled scrolls. In 2023, a team succeeded in reading substantial passages from one scroll. The text appears to belong to an Epicurean philosophical work discussing pleasure, scarcity, food, and sensory experience — words unread for nearly 2,000 years.

Landscapes That Vanished Underground

How AI is helping archaeologists reconstruct the past is not limited to objects you can hold. Entire landscapes can now be studied through LiDAR, satellite imagery, and computer-assisted analysis. LiDAR — Light Detection and Ranging — fires laser pulses from aircraft and maps the ground surface beneath vegetation. The resulting data is dense, complex, and often difficult to interpret without computational help.

In 2018, LiDAR survey results from northern Guatemala revealed a vast network of ancient Maya structures beneath the jungle canopy — more than 60,000 previously unidentified features, including causeways, reservoirs, houses, and defensive works. The scale was simply impossible to grasp through ground survey alone. Similar computational approaches are now being used to identify archaeological features in other landscapes, including medieval ridge-and-furrow field systems in England and possible buried sites in arid regions through satellite and radar imagery.

Where AI Hits Its Limits

Here is where the honest conversation has to happen. How AI is helping archaeologists reconstruct the past is a real story. But so is the risk of getting it wrong.

AI models learn from training data. When that data is thin — because a culture left few records, or because what survives is badly damaged — the model can produce plausible-looking errors. In epigraphy, for example, tools such as Ithaca can assist with restoring damaged Greek inscriptions, suggesting possible missing text, dates, and places of origin. But these are aids to scholarly judgement, not substitutes for it. A model may complete a pattern in a way that looks coherent while still lacking a firm evidential basis. The result can be persuasive, polished, and wrong.

The Manuscript Problem

Medieval manuscripts present a particular kind of challenge. Many survive only in fragments: leaves cut from books, pieces reused in bindings, and scraps separated from their original codices over centuries. It takes expertise to determine the origin, script, date, contents, and possible parent manuscript of a fragment. It is slow work.

New digital tools, including handwritten text recognition and machine-learning approaches to palaeography, can now narrow down some possibilities far more quickly than before. In Switzerland, the Fragmentarium project provides a digital research environment for medieval manuscript fragments, allowing scholars and institutions to catalogue, describe, transcribe, assemble, and compare disconnected leaves. Other projects are using machine learning to assist with transcription, script analysis, and the identification of scribal hands. These tools do not always get things right, but the leads they produce can be valuable for human scholars to test.

What Comes Next

The pace of development is fast. Over the past few years, AI and machine-learning tools have helped researchers read damaged texts, analyse vast bodies of imagery, identify archaeological features, and test reconstructions of broken or incomplete objects. Some of this work is genuinely transformative. Some of it remains experimental. Much of it depends on whether the data is strong enough, the method is transparent enough, and the human interpretation is careful enough.

For archaeologists, tools have always mattered. The discipline has long relied on trowels, photographs, drawings, measurements, typologies, maps, databases, and patient comparison. AI does not change the fundamental task. Archaeology is still the work of rebuilding the past from what survives: piece by piece, word by word, image by image, landscape by landscape. What AI may change is the speed, scale, and range of the questions archaeologists can ask — provided they never forget that a reconstruction is not the past itself, but an argument about the evidence.