
Most academic rules were written for a world where help came from a person. A model sits somewhere between a textbook, a tutor and a ghostwriter, and the difference between the first two and the last one is whether the thinking is yours. Here is where the line usually falls, and how to work with a model without risking your degree.
You may read explanations, ask questions, check your spelling, look up how a method works and discuss ideas. You may not submit text as your own work if you did not write it and reason through it. Two versions of similar help can land on opposite sides of that line: asking a model to explain how a confidence interval works is study, while asking it to write your results section is misconduct.
Students writing in a second language are often caught by the same trap. Improving grammar is normally allowed. Rewriting the whole argument into smoother prose so the ideas are no longer yours is not, even where the words started out as your own.
Writing practice questions and testing yourself. Asking for an explanation at a level below and above your current one. Asking what a sceptical examiner would attack in your reasoning so you can fix the gap. Getting a plain-language summary of a paper before you read it properly. Generating practice problems and marking your own attempts. Improving your English at sentence level, where your institution permits it. Reformatting references into the required style.
All of these keep the intellectual work with you. None of them produces a submitted artefact you did not compose.
Producing assignment text and pasting it in. Paraphrasing a source so that a model's wording stands in for your own reading. Writing a discussion post, a lab report section or a reflection. Using a model to answer questions while an exam is open. Submitting generated code you do not understand. Translating your essay during a language exam. These usually breach academic rules, and saying a tool was involved does not convert them into permitted help. Policies also differ between institutions and between modules, so a use that was fine last term can break the rule in this one.
Some of them are also just hard to get away with. A section written by a model does not match the voice, vocabulary or error patterns of your other work, and a marker who reads a hundred scripts a term notices that quickly.
Detection tools are unreliable, so do not assume they are the only safeguard. What gives a ghostwritten piece away is everything around it. You cannot discuss it in a tutorial. Your draft history is empty. The writing level jumps between paragraphs. Your reference list includes sources that do not exist because the model invented them. The version on a shared drive predates your first edit.
Most institutions now weight a short conversation about your own work far more heavily than a score, and that conversation is impossible to fake if the work is not yours. The risk is not a coin flip. It is slow, high-probability exposure.
Where a tool is allowed, disclose it the way you would disclose a conversation with a tutor. A line in a methods note or an acknowledgements section is usually enough: I used a language model to explain regression to the mean and to check my grammar; the argument and analysis are my own. Keep the chat log alongside your notes.
If your institution has no written policy, ask the module lead in writing before you submit. Being told to avoid a tool is a far better outcome than an academic misconduct hearing. Keep that reply, because a written instruction dated before submission is the strongest protection you can hold.