That's mostly a problem with GPT3, not GPT4. I'm not saying it doesn't make some of them up, but I've had great research experiences with it.

It's true that after you use the bot to fetch you the papers, you do still need to read them... but given what a dramatic difference there is between GPT3 and 4 I'd say this is a problem that will be utterly annihilated before most people even hear it exists.