my biggest issue with chat bots (specifically ChatGPT) is it will, with overwhelming confidence, assert it’s capable of doing something it isn’t. and it’ll often gaslight you in the process. I don’t think the simulated human approach works very well I expect this will be an interesting read
the titles and dates it gave in this example just seem to be generic titles it invented rather than say it couldn’t complete a task, which I find odd. bots have constraints, it shouldn’t have to pretend it can do something if it understandably can’t
I read the following paper yesterday "Evaluating Verifiability in Generative Search Engines?" Their finders are that GPTs need to be both fluent + informative AND have accurate + adequate citations. Many are simply fluent + informative, but not the latter. https://arxiv.org/abs/2304.09848