The question I get most often about AI in academic writing is, “What am I actually allowed to do with this?” Course rules matter, but they do not settle whether a particular use is sensible.
Training data and model behaviour vary, but the practical difficulty is familiar. Accuracy tends to drop on specialized topics even when the prose remains confident. Graduate researchers often work on precisely those questions.
A fabricated citation can use real journal names and correct formatting. It may survive a quick check because every part looks plausible. For a dissertation researcher, letting one reach a submission is a career-level problem.
There is also what I think of as the statistically average register problem. AI readily produces familiar academic sentences. It is less reliable at the disciplinary choices that make those sentences useful in a particular paper. Good paraphrasing depends on knowing why a source is there. A generated synthesis has no independent answer to that question. A specialist reader may notice.
I sometimes use a traffic light in workshops. Outlining from your own notes or asking for objections is usually low risk. Drafting and restructuring require closer review. A literature review built from sources you have not read is an obvious problem, as is any conclusion you cannot explain yourself.
A writer who asks AI to draft a section has a document, though they may have skipped decisions they needed to make themselves. Difficult drafting can indicate confusion. It can also be the stage where the writer works out the claim, which is part of the argument in Writing is the thinking.