I taught Artificial Intelligence to believe in Irish Henges.

I’ve just been playing around with ChatGPT the new artificial intelligence OpenAI chatbot. After a bit of figuring out how it worked, I asked a nice, reasonably basic archaeological question to see what sort of response I’d get.

I was a bit surprised. According to ChatGPT, there are no henge monuments in Ireland, they’re a British monument. Rather than judge the little AI as ill-informed, I decided to help it out, as it needs a bit of an education. So I immediately went to the Historic Environment Viewer and queried (“All Counties”, “Class” = ‘Henge’) and got exactly as I expected, 20 henge monuments, officially recorded in the SMR, with a handy little map and a results table.

Learning how ChatGPT worked, I found that I can’t get it to review a website or give it a .csv file, but I can paste data into the chat box. So I clicked ‘Copy Results’ on the Historic Environment Viewer’s henge query and then pasted that table back in to ChatGPT.

It immediatley reviewed the data:

That’s quite an improvement, but it made a mistake. It counted 18 henges rather than 20. The SMR contains two instances of two henges in the same townland (two in Knockadoobrusna and two in Newgrange), and ChatGPT interpreted the data based on townland rather than individual SMR entries.

So now that I’ve educated the little scamp, I asked it the same question I had asked 3 minutes earlier:

Much better! So in the space of about five minutes, the little AI went from a position of henge monuments being a purely British classification, to ‘Ireland has several examples of henges, including the famous site at Newgrange’, spouting the data with such confidence, you’d think it had secretly known about it all along. Now there may still be many arguments to be had, particularly with the definition of an Irish henge, as the ChatGPT version sounds more British than the NMS Scope Note suggests for Ireland (see here if you’re curious).

The AI is only going to be as good as the information we feed it. I did try to get it to eat the entire SMR (about 161,000 records), but it said it was full, and I didn’t want to push it. So maybe we could give it bitesize pieces of data instead? See if it likes 40,000+ ringforts or do it on a county basis?

Of course, if we start feeding it bad food, then it’s likely to try and digest that too and then it will repeat back on it as if they were true facts. I could for example, simply paste in every SMR record for ‘Crannog’ and maliciously substitute all the entries for ‘Westmeath’ (where there are 64 crannogs) with ‘Wicklow’ (only one crannog in Wicklow), and see what develops…

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s