How Is Artificial Intelligence Changing Archaeology?

POV Travel
Jul 15
8 min read

In the 1980s, a group of specialists examined a lump of carbonised papyrus from Herculaneum, damaged by earlier attempts to unroll it, and declared it unreadable.

Not difficult. Not awaiting better methods. Unreadable. The judgement of the best people in the field, working with the best tools available, was that whatever had been written on this object was gone forever and no one would ever recover it.

In June of 2026 the entire scroll was read. Twenty columns of Greek. Nobody touched it.

It was scanned with high energy X rays, unrolled inside a computer, and an artificial intelligence, trained to see the almost invisible difference between carbon ink and carbonised papyrus, picked the letters out of the noise. A text sealed for nearly two thousand years, and pronounced dead twice, is now being read by people who were not born when it was written off.

We would like you to hold that story next to every confident statement you have ever heard about what the ancient world can and cannot tell us.

How is artificial intelligence changing archaeology?

Quick Answer

In three ways. It reads texts that were physically unreadable, most spectacularly the carbonised Herculaneum scrolls, which are being unrolled and deciphered without being opened. It finds sites that human eyes have missed, scanning vast quantities of imagery. And it analyses data at a scale no human team could manage.

Roughly six hundred sealed scrolls remain. Parts of the villa that held them have never been excavated.

What it cannot do is tell us what any of it meant. And what it finds depends entirely on what it was trained to look for, which is a limitation nobody has solved.

Reading what could not be read

The Herculaneum library is the clearest demonstration, because the problem looked genuinely impossible.

In 79 AD, Vesuvius buried the town under thirty metres of ash. In one luxurious villa, a collection of perhaps eighteen hundred papyrus scrolls was carbonised in place. They survived, which is the miracle, and became unreadable, which is the tragedy. Discovered in 1752, they were sliced, pulled apart, destroyed by the score, and eventually abandoned in storerooms as an insoluble problem.

The difficulty was never simply that they were rolled up. It was the ink. Roman scribes wrote in carbon based ink on papyrus that the eruption turned to carbon. Scan the scroll and you get a three dimensional map of a charred cylinder in which the writing and the page are the same element. There is almost nothing to distinguish them.

The solution arrived from an unusual direction. A prize backed challenge, launched in 2023, invited anyone in the world to develop software capable of reading the scrolls. It worked in a way that closed academic projects often do not. A twenty one year old computer scientist found the first word. Substantial prize money was paid out. The techniques improved rapidly under open competition rather than slowly under peer review.

The text now being read appears to be a philosophical discussion of ethics, the arts and human behaviour, by an author nobody can yet name.

Around six hundred unopened scrolls remain in storage. Large sections of the villa have never been excavated. Scholars believe more may still be underground.

Pause on that. There is a Roman library sitting in a hillside, and we now possess a tool that can read books nobody can open. The lost literature of antiquity may not be lost. It may simply have been waiting for us to become capable of it.

Finding what nobody could see

The second transformation is happening across landscapes rather than libraries.

Archaeological survey has always been limited by human attention. Somebody must look at an aerial photograph, or a laser scan of a forest floor, and notice that a faint rectangular discolouration is a buried wall rather than a natural feature. It is slow, subjective, and there is far more landscape than there are archaeologists.

Machine learning inverts the arithmetic. Trained on known sites, a model scans thousands of square kilometres in hours, flagging anomalies for human checking. It does not tire at the four hundredth image.

In the deserts of Peru, researchers working with computing specialists trained systems to detect the figurative Nazca geoglyphs in aerial imagery. In roughly six months they identified more than three hundred previously unknown figures. Since the 1940s, patient human survey had found around four hundred and thirty. One project, in half a year, came close to doubling eighty years of work.

Similar approaches are detecting buried mounds and settlements from satellite data, including in Mesopotamia, where archaeologists are racing to identify sites from declassified Cold War spy satellite imagery before modern development erases them. The photographs were taken decades ago. Many of the landscapes they show have already gone. The machine is reading a record of a world that no longer exists.

The question nobody asks about the machine

Now the part that gets left out of the excited coverage, and it matters more than any of the above.

A model trained on known sites finds things that resemble known sites.

Read that again slowly. Every one of these systems learns from a catalogue assembled by previous archaeologists, working within previous assumptions about what a settlement looks like, what counts as a structure, which shapes are human and which are natural. The machine then goes out and finds more of the same.

Which means it may be systematically blind to precisely the thing that would teach us the most: the settlement that looks like nothing we have seen before. The site that does not conform. The structure built by people whose way of building we have not yet imagined.

Artificial intelligence in archaeology is, in this respect, an extraordinarily powerful engine for confirming what we already believe. It inherits the assumptions of its training data and it inherits them invisibly, because the model cannot explain itself and the people using it did not write the catalogue.

Nobody has solved this. It is not obvious that it can be solved.

There are other dangers too, and specialists have raised them. Every flagged anomaly still requires human verification, and enthusiasm runs ahead of confirmation. A dot on a map is not a discovery. And the same technology that lets archaeologists find undocumented sites from satellite imagery lets looters find them, and looters are not slowed down by permits, funding cycles or peer review. Detection is not protection.

What the machine cannot do at all

Artificial intelligence can find a buried structure. It cannot tell you what happened inside it. It can read the letters on a scroll. It cannot tell you what the writer believed, or feared, or was arguing against. It can detect three hundred geoglyphs in the Peruvian desert. It cannot tell you why anyone made an image on a scale too large to see from the ground.

The interpretive gap is not a technical problem awaiting a better model. It is the whole of archaeology. Meaning is not in the data. Meaning is what a human mind constructs when it holds the data alongside everything else it knows about being a person.

So the machine hands us more evidence, faster, and leaves the hard part exactly where it always was.

What we are really being told

There is a deeper thing happening here, and it is not about technology.

For most of the modern era, the story of the ancient world felt settled. The great discoveries had been made. The texts we had were the texts there were. The scrolls were unreadable. The record was closed. What remained was refinement at the margins.

Every part of that was wrong, and it was pronounced with enormous confidence by people who had every reason to know.

Six hundred sealed scrolls. Unexcavated rooms in a Roman villa. Hundreds of geoglyphs found in a single season. Buried Mesopotamian settlements identified from photographs taken by Cold War satellites. A thirty metre void inside the Great Pyramid, detected in 2017 by counting subatomic particles, in the most studied building on Earth, after two centuries of examination.

The past is not shrinking as we learn about it. It is getting larger. There will be more of it next year than there is now, and some of what arrives will not fit the story we currently tell.

We think that should change how you hold every confident account of the ancient world, including the ones you find in textbooks, and including this one. The experts who declared that scroll unreadable were not fools. They were simply working at the limits of what was then possible, and they mistook those limits for the limits of reality.

Everybody does this. It is what expertise feels like from the inside.

How POV Travel approaches this

We take travellers to ancient sites, and we tell them what is known, what is disputed, and what nobody has any idea about. That last category has always been the largest, and it is being redrawn right now by imaging physicists and computer scientists rather than by archaeologists.

This matters for how we talk to you. When we stand at Giza and say the precise engineering of the largest stones is still argued over, we mean it, and it is entirely possible that a scanning technique will settle it within your lifetime. When we say a text is lost, we now have to add a caveat, because lost is turning out to be a less permanent condition than it was.

What we will not do is pretend the machines have answered the questions. They have not. And we will not pretend the machines are neutral either, because a system trained on what we already found will keep finding what we already found.

We question. We teach. We leave you with an opinion of your own.

We would rather you understood that the story of the ancient world is being actively rewritten, this year, than that you thought of archaeology as a closed book with the answers printed in the back.

There is no back. There is only the next scroll, and nobody knows what is in it.

Frequently Asked Questions

How is AI used in archaeology?

Chiefly in three ways: reading damaged or sealed texts such as the carbonised Herculaneum scrolls, detecting buried sites in satellite and aerial imagery, and analysing large datasets including ancient DNA.

What are the Herculaneum scrolls?

A library of papyrus scrolls carbonised by the eruption of Vesuvius in 79 AD, discovered in 1752 and long considered unreadable. Scanning and machine learning have now allowed an entire scroll to be read without opening it, with hundreds more still sealed.

How did AI find new Nazca lines?

Researchers trained models to recognise figurative geoglyphs in aerial imagery. In roughly six months they identified more than three hundred previously unknown figures, close to doubling the number found by human survey since the 1940s.

What are the risks of using AI in archaeology?

Models trained on known sites may only find things resembling known sites, potentially missing the most unfamiliar and valuable discoveries. False positives require human checking, and the same tools help looters locate undocumented sites.

Can AI tell us what ancient sites meant?

No. It can locate and read. It cannot interpret. Meaning is constructed by human beings holding evidence alongside everything else they know, and no algorithm has any access to that.

Go and stand in the mystery

There is a difference between reading that a stone weighs a thousand tonnes and standing beneath it at Baalbek, or between reading about the King's Chamber and feeling it resonate around you. The deep past is stranger, and more physical, than any page can convey. We walk among the ancient sites whose age and scale still aren't fully explained, and leave you with questions of your own rather than answers handed down.

Explore the expeditions: Lost Civilisations & Ancient Sites →