Given how transfixed I am by Rings of Power season 2 (so much better than the original) I’ve been asking Perplexity background information about Tolkien lore to address my uncertainty about elements of the show e.g. if Sauron is a spirit then why does he turn into Venom-esque black goo when he dies? There’s no question too obscure for Perplexity to provide an answer, with ‘suggested questions’ rapidly spiralling off into hyper-obscurity as you go down a rabbit hole. It satisfies my curiosity without leading me to spend hours lost in fan wikis, even though it poses obvious question about the ethics of what generative search trained on those wikis will be doing to their visibility.
So I thought I’d test it in an area where I know the lure inside and out. In Jon Hickman’s epic Secret Wars there were lots of threads which I know were never fully elaborated. I’ve been asking Perplexity questions about this unresolved plot threads and it will consistently provide a singular and definitive answer for each one, even when they weren’t actually shown in the story arc. The problem as far as I can see isn’t hallucination in the classic GAI sense but rather stitching together partial inferences in sources access through retrieval-augmented generation.
There’s a hole in knowledge, part of a story that was gestured to but never actually told, which various people have filled in through more or less speculative means. Their speculative answers are drawn upon by Perplexity in order to provide a singular answer to a question which is in reality unanswerable. It resides as a unrealised creative intention in Jon Hickman’s mind rather than something out there in the world which can be veridically described. Yet Perplexity treats every question as having an answer, generating those answers in a way that papers over the fractures and gaps in the knowledge system.
The combination of GAI hallucination and the combinatorial dynamics of RAG is very interesting. I feel like I’ve not got the language to adequately describe this yet, but this was a first attempt to put this dynamic into words, because I believe it is inherent in generative search and will manifest in different ways in other RAG systems.
Mark Carrigan 