There's a whole cottage industry right now looking at the limits of #ChatGPT (e.g. it can't double-check claims or do external attribution) and turning those limits into sentences that begin "Language models can only ever ..." I get why this is a reassuring, widely circulated message, but I wish a few people circulating it would mention that there is already active research that seeks to address those limits. https://paperswithcode.com/paper/attributed-question-answering-evaluation-and