There is no single answer, and be suspicious of anyone giving one
Institutional positions on AI currently span the full range. Some universities forbid any generative use in assessed work. Others permit it for language support but not for content. Others actively teach it, requiring students to document their prompts as part of the submission. The same student can be within the rules in one module and in breach in the next, at the same institution.
That variability is the reason a general answer is not usable. The honest response to “is this cheating” is that it depends on rules written by your program, and that those rules are the ones that will be applied if the question ever becomes formal. A blog post — including this one — carries no weight in an academic misconduct hearing.
What generalizes is the underlying principle. Assessment is a claim about what you can do. Anything that leaves that claim accurate tends to be permitted; anything that makes it inaccurate tends not to be. Applying that principle will usually tell you where a specific use falls, and you can then check it against your actual policy.
Assistance, collaboration, and substitution
It helps to separate three different things people mean by “using AI.” The first is assistance with expression: fixing grammar, adjusting register, suggesting a clearer sentence for an idea that is already yours. This has a long, uncontroversial history — supervisors, writing centers, and professional proofreaders have always done it, and second-language writers have always needed more of it.
The second is structural collaboration: asking for an outline of material you have gathered, having a draft argument critiqued, being asked what a section is missing. This is where policies genuinely disagree. The reasoning behind it is your own, but the shape came from somewhere else, and programs that assess your ability to organize an argument may reasonably count that as assessed work.
The third is substitution: a model produces the claims, the analysis, the interpretation, or the sources, and you assemble the output. This is the case nearly every policy prohibits, and the objection does not depend on detection. Submitting reasoning you did not do misrepresents what you can do, which is what the assessment was measuring.
- Assistance — grammar, register, clarity on sentences whose content is yours.
- Collaboration — outlining, critique, structuring material you gathered. Policies vary most here.
- Substitution — claims, analysis, or sources you did not develop. Prohibited nearly everywhere.
Disclosure norms are still moving
Expectations around declaring AI use have shifted quickly and are still shifting. Several journals now require a statement in the methods or acknowledgements describing any generative tool used, while explicitly refusing to list such tools as authors, since an author must be able to take responsibility for the work. Many universities have adopted similar declaration requirements for theses.
Where disclosure is required, specificity is what makes it useful. “AI tools were used in the preparation of this manuscript” tells a reader nothing. “A language model was used to improve the phrasing of the introduction and discussion; all analysis, claims, and citations are the author’s” is a statement someone can actually evaluate, and it protects you rather than exposing you.
Where no policy exists, err toward disclosure to your supervisor. Norms are converging on transparency, and a use you disclosed and were told was fine is a settled matter, while an undisclosed use judged retrospectively under a new policy is not.
Detectors are not the arbiter, and they misfire
AI-detection tools deserve a clear-eyed description. They estimate the likelihood that text was machine-generated from statistical properties of the writing, and they are not reliable enough to settle a case on their own. Crucially, they flag writing by people whose first language is not English more often than writing by native speakers — plainer vocabulary and more regular sentence patterns look, to a detector, like machine output. That is a serious fairness problem, and it lands on students who have done nothing wrong.
Two things follow. First, a detector result is a reason to ask a question, not an answer to one; institutions that treat a score as a verdict are making an error that they will eventually have to correct. Second, and just as important, the response to an unreliable detector is not to try to write around it. Rewriting your prose to change a score is a losing game against tools that change monthly, and pursuing it confuses an integrity question with a technical one.
The answer that works is evidence. If you can show dated drafts, notes with sources attached, an outline that predates the prose, and a supervisor familiar with your work, you have something a detector score cannot contradict. Build that record as a matter of habit rather than assembling it in an emergency.
Questions worth asking yourself before you submit
A practical test: could you defend every claim in this document in a conversation, without notes? Could you explain why each source is there and what it says? Could you reproduce the reasoning that connects your results to your conclusions? If the answer to any of those is no, something in the document is not yours, regardless of how it got there.
A second test concerns the assessment itself. What is this task trying to find out about you? If it is measuring whether you can construct an argument, and something else constructed it, the result is misleading even where no rule names that specific case. If it is measuring whether you understand a method, and you understand it, help with the prose does not compromise the measurement.
None of this is an argument that these tools have no legitimate place in academic work. Used as a structured aid to work you are doing yourself, they are useful, and for writers working in a second language they can reduce a genuine and unfair disadvantage. The distinction that matters is between improving how you say something and outsourcing what you have to say.
