"I don't know" isn't in the training data. Nobody writes engineering books, science papers and blog posts that end with "well, I don't really know, the end".

Oh, that's interesting - good point, since it's filtered and not trained from scratch. My prior would be to assume it's just bs'ing as LLMs usually do but it seems worth exploring.