Study & Learning
Where Are the Actual Lines on Using AI for Coursework?
By Jim Vernon, Editor, AI Intelligence International · Published 18 August 2026 · Reviewed against our editorial standards · About the author
Most institutional AI policies say something like 'permitted for support but not for producing assessed work', which is genuinely ambiguous at the boundary where students actually operate.
Detection is unreliable in both directions, so the practical question is not what you can get away with but what standard you can defend if asked. This article gives one.
Key takeaways
- Detection tools are unreliable in both directions; do not treat a low score as safety or a high one as proof.
- The defensible test is whether you can reproduce and defend the work without the tool.
- Keep a process record — drafts, prompts, notes — for any substantial assessed piece.
- Read your specific course policy, because they differ sharply between modules.
Why is the policy usually ambiguous?
Because institutions are trying to permit legitimate support — explanation, feedback, proofreading — while prohibiting substitution, and there is no clean linguistic boundary between them.
The result is that policies name extremes and leave the middle to interpretation, while individual module leaders apply very different standards.
This means the only reliable source is your specific module's stated rules, in writing, which frequently differ from the university-level policy.
What test actually holds up?
Could you reproduce this work, or defend it in a viva, without the tool? If yes, the tool supported you. If no, it produced the work, whatever the policy language says.
This test maps closely to how academic misconduct panels reason in practice, and it survives the ambiguity of the written rules.
It also happens to protect your learning, which is the reason the rules exist.
What is usually clearly fine?
Explanation of concepts, generating practice questions, feedback on your own draft, grammar and clarity checking, and help understanding a source you have read.
Planning and structuring, provided the ideas and the argument are yours and you could justify the structure.
Translation support for non-native speakers, where most institutions are explicit that language support is permitted.
What is usually clearly not?
Submitting generated prose as your own writing, generating the argument or analysis that the assessment is measuring, and producing citations you have not read.
Fabricated references are the most commonly caught form and the most damaging, because they are objectively verifiable and indicate the source was never consulted.
Using it in a closed assessment, obviously, and in any context where the module states a prohibition.
What should you do about the ambiguous middle?
Ask the module leader in writing and keep the reply. A specific question — 'may I use it to critique my own draft?' — gets a specific answer, and the written record settles any later dispute.
Declare use where a declaration mechanism exists, in the terms the institution requests. Declared use is almost never penalised; undeclared equivalent use sometimes is.
When in doubt, do the version you could explain to a panel without discomfort. That instinct is a better guide than any policy reading.
How do you protect yourself against a false accusation?
Keep the process record: version history in your document editor, research notes, outline drafts, and dated files. This is by far the strongest defence available, and it costs nothing to maintain.
Write in an editor that preserves revision history rather than pasting a finished piece into a blank document, which produces a suspicious authoring trail even for entirely original work.
If accused, ask what evidence is being relied on, and offer your process record and a conversation about the content. Detector output alone is weak evidence and institutions increasingly acknowledge this.
Worked example: two dissertation approaches compared
Two final-year students on the same programme both used AI extensively on a 10,000-word dissertation, and were both asked about it at the viva stage after their supervisors flagged unusually polished prose.
The first had used it to draft sections from her notes, then edited heavily. She could not, under questioning, explain why she had chosen one methodological framing over another, because that choice had come from a draft she had accepted. The panel referred the case, and the outcome was a capped resubmission.
The second had used it for four things: generating critique of his own draft chapters, checking his statistical interpretation, practice questions to prepare for the viva, and proofreading. His literature review, argument and analysis were written from his own notes.
He also had a process record: 47 saved drafts with version history, a research log with 90 annotated sources, and a folder of prompts and responses he had kept from the start on his supervisor's advice.
At the viva he answered every question about his methodological choices directly, and offered the process record when asked. The matter was closed in the same session, and his declared use statement — four lines describing exactly what he had used and for what — was later used by the department as an example of good practice.
The distinguishing factor was not how much they used the tool. It was whether the assessed thinking was theirs.
Frequently asked questions
Do detectors actually work?
Not reliably. They produce both false positives, particularly for non-native English writers and for heavily edited formal prose, and false negatives on lightly rewritten output. Treat their output as a prompt for a conversation, not as evidence.
Should I declare use even if not required?
A brief, honest note is usually a net positive and costs nothing. It signals confidence and pre-empts suspicion about polished writing.
Is proofreading always allowed?
Usually, but language-focused assessments sometimes prohibit it because language is what is being marked. Check the specific rubric rather than assuming.
What if my institution has no policy at all?
Apply the reproduce-and-defend test and keep a process record. Absence of a policy is not permission, and standards are frequently applied retrospectively.