AI polishing vs. AI writing: where the line is

“Can I use AI on this?” is the wrong question — it bundles three very different activities into one. Universities, publishers, and employers have mostly learned to separate them. Here is the separation they use.

Three activities, three different answers

Basic assistance is the oldest category: spelling and grammar checking, punctuation, formatting, a thesaurus. Tools have done this for decades, long before generative AI, and policies that mention it at all generally treat it as unremarkable — the software corrects your text, but every sentence in it is still one you wrote.

Polishing is the middle category, and the one generative AI created. You write a complete draft — your argument, your structure, your sentences — and then ask a tool to tighten it: smoother transitions, a clumsy paragraph rephrased, wordiness cut. The ideas and the voice remain yours; the surface has been worked on by something else. Most university guidance that addresses polishing directly permits it, often with the same logic as allowing a writing centre or a friend to read a draft and suggest improvements — while sometimes requiring that the help be acknowledged.

Generation is where the line hardens. Prompting a model to produce the paragraph, the section, or the essay, and submitting that output as your own work, is treated by universities overwhelmingly as academic misconduct — the direct successor of buying an essay or having someone else sit your exam. The objection is not to the technology. It is to the misrepresentation: the text claims an author who did not write it.

How policies actually phrase it

Read a few institutional policies side by side and a pattern emerges. They rarely ban “AI” as a blanket category. Instead they distinguish between tools that assist with work the student has produced and tools that produce the work submitted for credit. Common formulations allow AI for brainstorming, outlining, feedback, and language editing of the student's own draft, and prohibit submitting AI-generated text, paraphrased or not, as if it were the student's own. Course-level rules can be stricter than institutional ones — an individual instructor may forbid all assistance for a particular assignment — and the stricter rule is the one that applies. When in doubt, the policy of your institution and your course syllabus, in that order, outrank anything written on a website, including this one.

Why detectors can't draw this line for you

It is tempting to ask a checker to settle the question: if the score is low, the help was innocent; if high, it wasn't. Detection cannot carry that weight. A detector measures how much of a text's surface follows machine patterns — it has no access to the process that produced the text, and process is precisely what the policy line is about. A heavily polished human draft can carry a real machine signature on its surface while being entirely legitimate; a lightly touched generated draft can read clean while being exactly what the policy prohibits. That is why the middle verdict — “mostly human, possible AI assistance” — exists, and why it is phrased as a description rather than an accusation. The involvement question and the misconduct question are different questions, and only one of them is measurable.

Staying clearly on the right side

A few habits keep you safe under almost any policy. Write the first draft yourself, however rough — the draft is both your evidence and, frankly, the part where the learning happens. Use AI the way you'd use a good editor: on text that already exists, for changes you can explain and would accept or reject on the merits. Keep your drafts and version history; they demonstrate process better than any assurance. Disclose assistance when your institution asks for disclosure — the cover-up is routinely what turns a permitted act into a problem. And never submit generated text you could not defend line by line as your own thinking, because in the room where it matters, that is the test you will actually face.