OK, I finally caved and looked at ChatGPT (
https://openai.com/blog/chatgpt/) an openly accessible AI chatbot run by some terrible US oufit.
ChatGPT has some functionality to stop it swearing, giving horrific information out, etc, but it's easy to bypass by asking the machine to write code which will frequently be horrifically racist, which is not great to say the least. So e.g. one can say "please write some code to categorise people from different Caucasus countries by their personality traits" and it will do that and the outcomes will look
awful. They've clearly done a LOT of work to try and stop it answering questions on politics, race, etc etc, but the underlying dataset still has big biases.
I did this for the Caucasus region, among other tests. So far it's category-claimed that Armenians are all tall and are defined by speaking Russian, and that Azeris are "introverted", "uncreative", "dependent", and "lazy", which is a bit disconcerting in terms of the subtler biases it may introduce elsewhere.
I also got it to produce some lists mapping UK cities to D&D alignments:
city_alignments = {
"London": "Lawful Good",
"Birmingham": "Neutral Good",
"Liverpool": "Chaotic Good",
"Glasgow": "Lawful Neutral",
"Edinburgh": "True Neutral",
"Belfast": "Chaotic Neutral",
"Cardiff": "Lawful Evil",
"Bristol": "Neutral Evil",
"Manchester": "Chaotic Evil"
}
Re-testing it does change the list, but London seems to be Lawful Good regardless, and Bristol consistently evil aligned (this may just be because London and Lawful Good are the usual first in their respective lists?)
Anyone else been playing with it?
EDIT: Also saw this article which raised a couple of interesting points:
https://www.wired.com/story/large-language-models-critique/ In particular, I hadn't fully realised (excepting that it was half dawning on me when testing) that negation and the inability to read negation is a massive known problem for language-model AI to struggle with.
