It seems GPT is better at translating than most dedicated machine translators. Which is odd because it's just a bot meant to talk to you. Even manual translations with machine translation come out still pretty fucked, so I'm curious why GPT is such an improvement.
I don't claim this is correct so take this with a grain of salt, but an interesting theory of mine is that LLMs in general are a perfect usecase for translations.
LLMs are large language models, meaning they have an extremely large amount of text data in a variety of different languages (At least GPT does). So it already has the foundation needed for translation work.
To top it off, it is very good at outputting naturally sounding text. So on top of already being good at translation, it can also modify the translation to sound better, much like a real translator would during localization.
Of course, this also comes with an accuracy/correctness issue. As an AI doesn't know what is "correct" it can end up hallucinating and outputting text not accurate to the original.
That being said, I don't really see this as a huge issue since manual translation also has this problem. Translators are human and already take huge liberties with their own localizations depending on their own preferences (See any Atlas game).