Well, according to it, the design was to always give users an answer, and that sometimes results in made up answers like that. Super comforting, but at least it was honest. Seems like something a company wouldn't want people to know, but this stuff isn't the easiest to predict. It made up authors and paper titles that I couldn't find any evidence for. It also included links, but to completely different papers, if they resolved.

I don't think that's a user or training data issue, it's probably exactly what it said it was.