They’re great at proofreading. They’re also good at writing conclusions and abstracts for articles, which is basically synthesising the results of the article and making it sexy (a task most scientists are hopelessly terrible at). With caveats:

- all the information needs to be in the prompt, or they will hallucinate;

- the result is not good enough to submit without some re-writing, but more than enough to get started and iterate instead of staring at a blank screen.

I want to use them to write methods sections, because that is basically the exact same information repeated in every article, but the actual sentences need to be different each time. But so far I don’t trust them to be accurate with technical details. They’re language models, they have no knowledge or understanding.