> Therefore, our safeguard is constrained by time, knowledge, and resources. The problem then is that if that safeguard fails and A-Grade Bullshit is taken by an author using the AI as a research assistant, put into a research paper, book chapter, or online article, and is then cited. That becomes potential training data for the AI if it is released into the world or has the window of its training data set extended.

I don't see the value in GPT models for this kind of use. I see the value in like, translation work, or summaries and other transformations, and that GPT works best when in concert with search engines, document embeddings, or knowledge databases like Wolfram Alpha. It's good at predicting more text which is useful in many different kinds of contexts but not in all contexts. It's pretty clear that it is untrustworthy for factual responses like quotes and population counts but that problem goes away if the actual factual data is included in the prompt!