@arstechnica
This isn't going away with cleaner data. The problem is the models inherent lack of ability to tag jokes and fiction. Again these models are essentially predictive text. LLMs were never meant to return factual information and are not programmed to verify their answers.