Bold of you to assume they’re curating their training data.

Those underpaid Indians probably aren’t very good at picking up irony, even if they give a shit.

Most of the curation or fine tuning is done in low income African countries so this is little surprising. They‘re cheap labour but you can‘t expect them to reliably detect sarcasm or notice mistakes in specialized fields. They basically give a thumbs up whenever the AI sounds convincing. Of course that includes instances where it‘s confidently wrong and that appears to be most of the time with this model.

It’s not a training data issue, look up Retrieval Augmented Generation. It’s basically serving up stuff on the web and taking it as gospel.