AI in Archaeology: Real Discoveries or Hype?

Yes, AI is genuinely discovering things in archaeology, but not on its own. It scans very large datasets, such as lidar and satellite imagery, far faster than manual survey, and it has made it possible to read carbonised scrolls without unrolling them. In every case reviewed here, archaeologists checked the output, and false positives remain a known problem.

The records are strongest for three projects: the reading of the Herculaneum scrolls, a survey of the Nazca geoglyphs and a deep-learning study of Maya lidar. This article sets out what each one shows, what rests only on press reporting, and which popular claims could not be confirmed. For a general method of testing surprising stories, see a practical checklist for evaluating a claim.

The short answer: yes, but people still check the results

The useful work is mostly unglamorous. Lidar and satellite surveys produce more data than teams can comfortably inspect by eye, and machine-learning models can flag places worth a closer look. That turns a task that once took years into something much quicker, though the flagged places are candidates, not conclusions.

The second achievement is different in kind. Machine learning has helped reveal writing inside papyrus scrolls that were never meant to be opened again. Here too, specialists interpret and translate the result.

No case turned up in this research in which AI alone established an archaeological site without human verification on the ground. The strongest documented evidence sits in the scroll reading, the Nazca survey and the Maya lidar study.

Reading scrolls that were never meant to be unrolled

The Villa of the Papyri at Herculaneum held about 1,000 papyrus scrolls. The eruption of Vesuvius carbonised them rather than destroying them outright, which left them as fragile black cylinders. Brent Seales, a computer scientist at the University of Kentucky, began work on non-invasive reading in 2007 and spent about 15 years on it. He succeeded with other scrolls, but the Herculaneum papyri resisted.

The breakthrough came through an open-source challenge, begun in March 2023, with prizes including one of US$1 million. More than 1,000 teams have entered. The method has three stages: a high-resolution scan using a particle accelerator, a virtual flattening of the scroll's structure, and machine-learning models trained to detect carbon ink against the carbonised background.

The first Greek letters were detected in October 2023. In February 2024 the prize winners, Youssef Nader, Luke Farritor and Julian Schilliger, were announced. Their model revealed parts of 15 columns from the innermost part of one scroll, probably a text on ethics by Philodemus.

A later reading of scroll PHerc. 1667 was carried out in full. The scroll was imaged at the European Synchrotron in Grenoble, and papyrologists translated about 22 columns of Greek, roughly 1.4 metres of writing, in what is believed to be a Stoic treatise on ethics. The reading is described as the first Herculaneum scroll recovered in full by these methods. Parts of over 300 scrolls in Naples are still waiting to be read.

The Nazca lines and a reported 303 new geoglyphs

A team from Yamagata University's Nazca Institute and IBM reportedly identified 303 new geoglyphs on Peru's coastal desert, a number reports say roughly doubles the known Nazca lines. The work was said to have taken six months and to have been published in PNAS. The underlying paper was not checked directly, so these figures rest on press reporting.

The designs were etched between about 200 BC and AD 650 and include animals, plants and geometric patterns. The method trained AI on thousands of aerial images and processed satellite and drone imagery. Archaeologists then confirmed every flagged site on the ground. Many of the new figures are interpreted as ceremonial paths, a reading linked to the earlier work of Maria Reiche.

Some researchers quoted in coverage stress that the technology is not perfect. Dr Alexandra Karamitrou of the University of Southampton warned that AI has limits in this field, and false positives remain a problem.

Mapping Maya settlements under the forest canopy

Dense jungle hindered early Maya mapping, which relied on ground survey with tapes and transits. Airborne lidar changed that, because its pulses can penetrate the forest canopy and reveal features on the ground beneath.

A study in the Journal of Archaeological Method and Theory, published online in November 2025, took the next step. Its authors, including Benjamin Britton, Nicholas Dunning and Lin Liu, built a multi-region convolutional neural network called Q2000, trained on data across about 35,584 km² of the Maya area. The model reached an F1 score of 0.89, with accuracy comparable to earlier local studies, even with a relatively small training sample.

The study uses a human-in-the-loop framework, meaning expert knowledge is built into the workflow rather than replaced by it. The same principle of human checking runs through all three projects.

Claims that are not yet proven

A widely repeated figure of more than 60,000 previously unknown Maya structures could not be confirmed from a primary source, and the Maya paper read here does not give it. It should not be treated as established. Likewise, some claims about a more recent reading of PHerc. 1667, describing an end-to-end recovery and independent confirmation of another scroll, come from a third-party blog and are not verified here.

Other AI applications appear in search results, including robotics at Pompeii, shell rings on the Atlantic coast and a Michigan caribou model. These were not checked against original pages, so they are left aside.

The pattern is familiar from other topics on this site. A striking claim travels faster than its evidence, as with the ancient aliens theory about the pyramids. Archaeology rewards patience with the record.

How to judge the next AI discovery headline

First, ask whether the claim describes a candidate site or a confirmed one. AI typically flags candidates that people then check. Second, look for a peer-reviewed paper, a stated dataset size and any reported error rate or false positive figure. The Maya study gives both its area and its F1 score, which is the kind of detail to look for.

Be cautious with round numbers and with figures that appear only in blogs or aggregated snippets. Then compare the claim with what the original research says, and with how the wider field has received it. That is roughly how findings move towards scientific consensus, through checking and repeated testing, not a single announcement.

Modern analysis of old objects can be impressive without being mysterious, as the Antikythera Mechanism shows. The same applies here: the tools are new, but the standard of proof is not.