What academic research is ChatGPT accessing?

"QUESTION: I don't know if this is a stupid one. Does LLM in the form of ChatGPT use information in research papers that are behind a paywall? If yes, what is the technology going on there to provide access? If no, is this a huge argument for OA? It has triggered a pretty robust discussion in Twitter (surprisingly, I thought everyone had left). Because of the tendril-like nature of Twitter discussions I am pulling the salient points together here because some people (including myself) might like to look into some of these references. The discussion broke into a series of questions (elaborated below): Do these AI systems have access to paywalled content? How well can ChatGPT actually ‘read’ the internet? Is this an argument for open access to research output? Does research output even matter given its small proportion of the internet? Should we be worried we don't know? What are the bigger implications here? What’s the take home?..."