Forget what you’ve heard about language models and AI, or what you already think you know, ok?
First of all, archaeology has been using machine learning for years; up until quite recently, you could say ‘ai’ and be understood that you were talking about things like agent based modeling or computer vision analysis and so on.
There is no such thing as artificial general intelligence.
The version of ‘ai’ that uses neural networks to represent statistical associations between words and then welds a chat bot on top of it to create the illusion of intelligence has a lot of marketing dollars behind it. For it to be financially successful, it doesn’t have to actually be intelligent or do any of the things its marketers say it can do. It just has to convince a manager that the appearance of doing the job is the same as actually doing the job. It confuses the appearance of an image with the process of doing art (and in so doing, participates in the aesthetics of fascism and the devaluation of expertise).
You can even make one in the physical world with slips of paper and a matchbox:
As representations of language I find these things interesting. Text can be meaningful even in the absence of authorial intention. For some things, like code, the associations are strong enough that yes useful code can be generated from a description of what is desired. But that’s not the same thing as just accepting whatever gets spat out. If you have deep experience of code, careful nuance and iterative interactions can help accelerate some aspects of the work; I direct you to people like Drew Breunig and Simon Willison. You should also pay attenion to Erky Salvaggio and Emily Bender’s group.
But you really need to know: it’s models all the way down and you are dealing with systems of systems all bolted together.
For me, then, ‘ai’ in archaeology are the models and machine learning approaches where we understand what’s going on under the hood. Generative ai on the other hand is an impenetrable web of statistical associations and unpredictable outputs.
Seance Time
Because these are models of language, models of culture, we’re going to use one to surface latent patterns in the published writings of a long-dead archaeologist. The approach we’ll use is similar to what Chantal Brousseau used for exploring Brexit discourses. Fire up your workbench and in the week seven folder you’ll find the notebook practical-necromancy.ipynb.
This is computationally heavy, and for some of you, it might not be possible for your machine to run it. You can use the Google colab service instead. You’ll need a gmail account, of course. Click the link, start a new notebook, then copy and paste each cell from the practical necromancy computational notebook into Colab. Click ‘runtime -> change runtime type’ and select ‘T4 GPU.’ Then run each cell from top to bottom in turn.
Make observations on what you see. How is this different than what you’ve come to expect about AI? What difference does making a ‘completion’ make to your perception of the model’s ‘intelligence’? What archaeological discourses seem to surface in Petrie’s writings? Is the model just repeating verbatim something he wrote? How would you know? Is there any value in reflecting archaeological materials back at you this way?
Archaeologists are exploring generative AI, by the way. How could we not? These are still human artefacts that tell us something about humanity. But so far, most published interventions are more in the vein of ‘how can we use this’ rather than ’this is a model of culture what does it tell us’, eg:
Ciccone, Gabriele. 2024. “ChatGPT as a Digital Assistant for Archaeology: Insights from the Smart Anomaly Detection Assistant Development” Heritage 7, no. 10: 5428-5445. https://doi.org/10.3390/heritage7100256
Graham S, Yates D and El-Roby A. Investigating antiquities trafficking with generative pre-trained transformer (GPT)-3 enabled knowledge graphs: A case study [version 1; peer review: 2 approved]. Open Res Europe 2023, 3:100 (https://doi.org/10.12688/openreseurope.16003.1)
Finally, here is an application that uses the excellent potential of neural networks/llm/‘ai’ to deal with the formulaic (ie, predictable) expressions of Latin epigraphy, and is worth exploring: https://predictingthepast.com/aeneas.