AI is not what you think it is

Or maybe it is, but it certainly isn’t what I thought it was, a year or so ago. There’s been a lot of attention lately on tools like ChatGPT and Copilot, systems that generate text, answer questions, and increasingly have found their way into academic life. These are examples of what are termed large language models (LLMs), and the term AI in everyday use is usually referring to these LLMs. But as I have come to realise, they represent just one branch of ‘artificial intelligence'. AI has long played a quieter role to help researchers make sense of the vast and complex datasets produced by advanced imaging techniques, environmental modelling, and material analysis. 

Since joining Durham in April this year, and working to develop applications of XR-CT (X-ray computed tomography) for heritage science, I’ve become increasingly aware of just how crucial AI is for making sense of this kind of data. XR-CT lets us look inside objects such as pottery, bones, sediments, without breaking them open, producing incredibly detailed 3D images of these objects and their internal structures. But these scans generate huge datasets, made up of thousands of ‘slices’ or X-ray images, each packed with subtle variations in density and texture (a single sample scan could easily produce hundreds of GB). These individual slices are combined (the computed tomography bit), to produce the 3D reconstruction of an object.

One of the biggest challenges with XR-CT data is something called segmentation. This is the process of telling the computer what it is we are looking at, usually involving drawing boundaries between different materials or features in a scan. This needs to be done on the individual X-ray slices, and there can be thousands of these in a single scan. Imagine slicing through a loaf of fruitcake and trying to digitally separate the raisins, cherries, and sponge. That’s segmentation. In archaeological terms, we might be trying to distinguish bone from sediment, or ceramic from organic residue. But it’s rarely that straightforward. Materials blur, densities overlap, and the human eye can't always distinguish between them. Manually picking through thousands of slices can be a real needle in a haystack situation, and you can imagine that the time it would take to do this manually is simply not feasible. This is where AI comes in. AI and machine learning algorithms can be trained to spot these differences in materials automatically, learning from examples and improving over time, and can spot patterns we’d never notice.

Another example is working with remote sensing data. My colleague Dr Louise Rayne, for example, develops bespoke ‘change detection’ algorithms for remote sensing data. These are tools that can automatically spot differences between scans taken at different times or under different conditions. She is developing these tools to automatically map and predict desertification in arid landscapes, detecting subtle shifts in vegetation and soil structure that would be almost impossible to track by eye. 

So what exactly is change detection? Imagine you’ve scanned an important archaeological artefact using XR-CT, and then you scan it again after a few months of environmental exposure, or after a conservation treatment. To the human eye, the two scans might look almost identical. But AI can compare them pixel by pixel, slice by slice, and flag up even the tiniest shifts, like a crack forming, a mineral dissolving, or a void expanding. Or at the landscape scale, a satellite has taken snapshots of a region on a daily basis, as the environment changes seasonally, annually, over decades and even longer. Louise’s algorithms are trained to spot these changes automatically, which means we can track deterioration, monitor conservation efforts, and mitigate problems before they become too widespread.

What fascinates me most is the sheer potential of these AI tools, not just for speeding things up, but for revealing things we might never have seen otherwise. Take thin section sediment micromorphology, for example: a technique I’ve loved since I first started peering down a microscope. Micromorphology produces incredibly rich data, and I truly believe that it is crucial for understanding formation processes and context, and thus informing how we interpret all other archaeological materials. It’s incredibly powerful, but also incredibly, painstakingly slow. Each slide can take hours, if not days or weeks, to describe and interpret.  AI could change that. Imagine being able to train a model to spot subtle signals in thin section images, features that the human eye might miss, or that we might not even know to look for. It’s not about replacing expertise but enhancing it. Tools that help us interpret micromorphology data more quickly and consistently could finally bring this technique to the forefront (where I think it should be), allowing it to contribute in real time to broader archaeological questions, rather than lagging behind while other methods deliver their “answers” faster.

If we want AI to help us here, we first need to digitise these collections and make them available. At Durham, our Digital Visualisation Lab in Archaeology has been working on just that. Our technical expert Alexander Jansen is using a Phase One IQ4 150MP camera to capture ultra-high-resolution images of thin section slides under both plane-polarised and cross-polarised light. These images, along with detailed metadata, could form the foundation for training machine learning models that could help identify features, classify materials, and even suggest interpretations. Once digitised, these slides will become part of a living dataset, ready to be explored, compared, and reinterpreted.

So now we know that AI is more than just ChatGPT. But actually the underlying principles between these tools are similar. Large language models like ChatGPT are trained on vast amounts of text, learning patterns in language so they can predict what word or sentence might come next. LLMs are doing something very similar to what we ask of AI in XR-CT segmentation or remote sensing: they’re finding patterns in complex data. In CT scans, segmentation algorithms learn to distinguish between different materials based on subtle differences in pixel intensity. In remote sensing, change detection algorithms track shifts in landscape features over time. All of these tools rely on training data, pattern recognition, and statistical modelling. They’re not infallible (and one of the key roles for domain experts is in the 'ground truthing' to ensure the predictions are accurate), but when used carefully, they can help us see more clearly, work more efficiently, and ask better questions. 

The key thing to remember is that all models have their limitations. Whether it’s a chatbot or a change detection algorithm, the key is knowing what the tool is good at, and what it’s not. LLMs like ChatGPT don’t ‘understand’ language in the way humans do. They don’t have knowledge or opinions; they generate text based on statistical patterns in the data they were trained on. That means they can sound convincing while being completely wrong! They’re also sensitive to how questions are phrased, and can struggle with ambiguity, nuance, or anything that requires genuine reasoning. Unlike segmentation algorithms trained on CT scans or change detection models in remote sensing, which are grounded in physical data, LLMs work with words, and words are fluid, context-dependent, and can carry layers of meaning. Even humans find the interpretation of words challenging. But I think this is also why I find LLMs fascinating; we can think about AI outputs as a form of material culture, reflecting ourselves back at us. They’re brilliant for brainstorming, summarising, and drafting, but they’re not a substitute for critical thinking or domain expertise (look at me, using the word ‘domain’, did you know that in software engineering, a domain refers to the targeted subject area of a computer program?). Used wisely, they can save time and spark ideas. Used uncritically, they can mislead. The trick is knowing when to trust them, and when to double-check.

Since moving to Durham, I’ve found myself increasingly interested in the applications of computer science for archaeology and heritage science. It’s been a genuinely fortuitous development. Researchers like Dr Amir Atapour-Arbarghouei and Dr Stuart James have long-standing interests in CT data analysis and computer vision, and have been very open to my ideas; their work opens up so many exciting possibilities for collaboration. I am excited by how much common ground there is, whether it’s segmenting CT scans of archaeological materials, detecting landscape change from satellite imagery, or training models to interpret thin section slides. The tools that could be the future of heritage science are part of a wider ecosystem of AI research, and being able to draw on that expertise locally has been transformative. My main frustration right now is that I can see the potential of these tools, and how they could transform the way we work with archaeological and heritage materials, but I don’t yet have the technical knowledge to use them fully. I’m looking at ways to upskill, to understand how these models and algorithms work, so I can apply them effectively and responsibly in my own research. It’s going to be a steep learning curve, but an exciting one, so watch this space!