"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".

A bit like publishing bias where only positive results end up in scientific papers.