When I was refining GPT-2 to make fake tweets of Donald Trump I encountered a similar problem about the DOI problem (and it would apply for anyone posting tweets with URLs). The twitter links would look legit because GPT-2 tried its best to create legitimate links. Here DOIs aren't being filtered by GPT3.5 so it does its best to generate a DOI that will look legitimate to a human to be correct but it will link to nothingness. So basically this is noise that GPT-3.5 needs to filter out.