I recently used ChatGPT (the free 3.5 Turbo version) to list five articles about public health surveillance, asking for ones published since 2017. The list jus

I recently used ChatGPT (the free 3.5 Turbo version) to list five articles about public health surveillance, asking for ones published since 2017. The list just had the article titles, authors, publications, and years. I had trouble finding the first one, so I asked the model for DOI numbers. It happily restated the article titles with DOI numbers. None of the articles were real. The authors were real (and actively published work after 2017). The publications were real, and often featured the listed authors. But the articles were all fake. Only one of the DOI numbers was legit, but it pointed to a different article (partial credit for matching the listed publication). So it can easily hallucinate not just nitty gritty details, but basic information. Thinking about it, the GPT models are trained on output, so they've picked up grammar, syntax, and conceptual relations between tokens (e.g., "eagle" is closer to "worm" than it is to "elegance"). But there are few, if any, examples of "first draft vs final draft." That could've been useful to pick up on how discretion and corrections are used in writing.