The question I get most often about AI in academic writing isn’t about plagiarism policies or detection software. The question is: “what am I actually allowed to do with this?” My answer is that the policy question is less interesting than the cognitive one.
Large language models are trained mostly on open-access English text, skewed toward the global north and recent publications. That will probably change. The deeper problem remains: accuracy drops as topics get more specialized, and confidence often does not. Bad combination for graduate researchers working at the edges of their fields.
A model can produce a citation that looks exactly right. Plausible authors. Real journal names. Correct formatting. Source does not exist. For a dissertation researcher, a fabricated citation that survives to submission is a career-level problem.
There’s also what I think of as the statistically average register problem. AI defaults to the most common academic prose patterns, which strips out the disciplinary specificity and productive complexity that distinguishes good scholarly writing from text that merely passes at a glance. Good paraphrasing requires knowing why you’re using a source. AI-generated synthesis doesn’t know that, and it shows. Well, it shows for experts in certain fields anyway.
The framework I use in workshops is a traffic light, which is common enough to be useful. Green uses: outline from notes, plain-language translation, pressure-testing your reasoning. Yellow uses: drafting from an outline you wrote, restructuring prose for flow. Red uses: literature reviews from unread sources, conclusions you cannot explain, output you cannot defend.
The question worth holding onto: is the AI thinking for me?
It matters because Writing is the thinking. A writer who asks AI to draft a section has a document. They may not have the understanding that comes from working through the ideas on the page. The discomfort of difficult drafting is not always a problem to solve. Sometimes it is the work happening.