With a bit more refinement, if it had the ability to generate graphs, etc, it might be able to generate very believable papers. At least, believable enough that you can't tell without reading the paper attentively.
Instead of spending 6 months laboriously doing experiments in a lab, and then a month writing up their results in a paper, researchers can already write a paper in one month if they just invent the numbers without actually doing the experiments. Peer review doesn't check for this. This only further reduces that one month to 5 minutes (+ hours of fiddling with LaTeX templates?). But in both cases if it gets found out your career is over...
However what about generating patents? To get a patent you don't need to have done any experiments that prove your technique actually works :)
Everything has an API. You can give it your data (or ask it to create some), and then ask it to write LaTeX, D3, MermaidJS, or code from any other framework that creates graphics. Problem solved.
If the thing you want to use is fairly popular and published on the web, even recently, it probably knows how to use it and combine it with everything else it knows.
Just yesterday I asked it to combine LiveView with a third party JS library to build an interactive thing and it got it on the first try using the latest Phoenix 1.7 RC which only came out in November.
I haven’t tried it, but I bet you could ask it to generate a PDF in code using your favorite language with text from GPT and graphics generated from any framework that’s compatible with your language. White-paper-as-a-Service.
