Following the advice of a local language learner, I’ve been experimenting with texting ChatGPT in Welsh. The results so far seem pretty good, until I asked it to generate a list of pedwar perlysiau (4 herbs), and it gave me 4 vegetables. The listing for carrots had a couple things that seemed strange to me.
“1. Moron (neu môr-lysiau) - Mae moron yn cael eu hoffi gan lawer o bobl am eu blas llwyd hynod o fywiog. Gallwch eu tyfu yn eich ardd neu mewn gweithdai tyfu.”
So 2 questions:
Is môr-lysiau actually used to refer to carrots? From what I could tell from searching, it means seaweed.
If I read this right, it’s saying that carrots have a ‘grey flavor’ (blas llwyd.) Does that make sense to anyone else, or is the AI just having a bit of fun there?
No, no-one would use this for ‘carrots’, it is, as you’ve found, ‘seaweed’.
Yes, that would translate as a grey taste - llwyd can also sometimes be interpreted as unremarkable, uninteresting, and insignificant, and also mouldy or musty! I’ve never heard llwyd applied to taste before, but that’s not to say that someone, somewhere, doesn’t use it.
I would trust ChatGPT less than Google Translate to be honest!
This is a good example of how ChatGPT works – it’s a giant predictive text machine, and isn’t at all intelligent! It can’t understand the meaning of anything, it can only look at patterns in its training data and try to reproduce them in context. I think it’s an interesting practice tool, but you do have to be very careful not to take anything it says at face value!
Definitely. It has been fun to play with a bit, but I’ve been double checking everything it says with dictionaries and a few different sources. It’s ok for a bit of fun, but ultimately I think typing to another Welsh learner would be better - because even if they’re making mistakes, I think it’s usually understandable where they are coming from.
I also tried getting it to generate some rude insults so I could make practice sentences a little more fun (becuase I usually like to pretend whatever simple sentences I’m learning are from people having a heated argument - because then it’s just so silly to be all about “would you like a red apple? No, absolutely not.”) - and it resisted several times because the output is trained to be pretty conservative with anything like that. I finally got it to output something. It was just clutching pearls trying to make sure I understood that I should never, ever consider saying such horrible things to an actual living person - but the actual result was just something mild like, “Stop talking to me I don’t like you”
I started taking Spanish in Middle school (equivalent to year 7 - 9), and our teacher would never have taught us any swearing, so that is definitely the first thing I would have tried then was to learn all the naughty words; so after finding out it just outright refuses, I’m pretty done with chat GPT as a language practice partner.