Use AI for Structure and Stuck Points, Not for Prose
There is a version of "using AI for writing" that is genuinely defensible, widely useful, and unlikely to get anyone into trouble, and there is a version that quietly hollows out the whole point of writing something yourself. The dividing line is not how much you use the tool; it is what you use it for. Structure and stuck points are where AI assistance holds up under scrutiny. Prose — the actual sentences that end up in your document — is where it stops being a tool you are using and starts being a substitute for the thinking the assignment was supposed to require.
What "structure, not prose" means in practice
Structure is the architecture of an argument before it has sentences: what claims you are making, in what order, with what evidence attached to each one, and where the piece needs a transition, a counterargument, or a worked example. An AI assistant is genuinely good at proposing a first-pass structure — three possible orderings for your argument, a checklist of sections a paper like this typically needs, or a list of objections a skeptical reader would raise against your thesis. None of that output is prose you would ever paste into your document. It is scaffolding you react to, rearrange, and then build your own sentences around. The same applies to stuck points: when a transition will not come, or a concept genuinely confuses you, describing the problem to a model and asking for a plain-language explanation or three ways to bridge two ideas can unstick you in minutes. You still write the transition. You still explain the concept in your own words once you understand it. The model got you moving; it did not do the moving for you.
Why this boundary holds up better than a cost-based one
It is tempting to think a boundary like this is really about not wanting to pay for a longer AI response, but the actual token-cost arithmetic shows that reasoning does not hold. Take a realistic outline-only exchange: a 150-word description of your assignment as input, and a 300-word structural outline as output. At placeholder rates of $3 per million input tokens and $15 per million output tokens, that comes out to roughly $0.0066 — two-thirds of a cent. Now take the alternative some people are tempted by: the same 150-word input, but asking for a full 2,500-word draft instead of an outline. The same placeholder rates put that at about $0.0506 — a bit over five cents. That is roughly seven or eight times more expensive in raw token terms, and yet it is still, in absolute dollars, close to nothing. If the only thing stopping someone from having AI draft an entire paper were the price, the arithmetic above shows that price is not much of a deterrent at all.
That is exactly why cost was never the real argument against it. The reason to stop at structure is not that a full draft is expensive — it demonstrably is not — but that a document you did not think through yourself is not something you can honestly claim, defend under questioning, or trust to be accurate, regardless of how cheaply it was produced. Framing the boundary around integrity rather than price also survives the inevitable future where prices keep falling: a boundary based on "it costs too much" erodes every time rates drop, while a boundary based on "I need to have actually done the thinking" does not move at all.
The pressure-test pattern
One of the most useful structural uses of AI is adversarial rather than generative: instead of asking it to produce something, ask it to attack something you already produced. Give it your thesis and ask for the strongest objections a skeptical reader would raise. Give it your outline and ask what a domain expert would notice is missing. Give it a paragraph you are proud of and ask what a careful editor would flag as unsupported. In every one of these uses, the model is not writing your document — it is functioning as a tireless, occasionally sharp reviewer who never gets tired of reading one more draft, and the actual response to its objections is still entirely something you have to think through and write yourself.
Recognizing when you have crossed from structure into prose
The test is not complicated, but it is worth stating explicitly because the line can blur mid-session: if text generated by the tool is going to appear in your submitted document substantially as written, you have crossed from structure into prose, no matter how you got there. This includes the subtler versions of the mistake — asking for "just a rough paragraph to get started" and then lightly editing it rather than replacing it, or asking the model to "make this flow better" and accepting a rewrite wholesale rather than identifying what specifically was unclear and fixing it yourself. Both of those feel like editing assistance in the moment, but if what survives into the final document is mostly the model's sentences with your light touch-up, the sentences are not yours, and you should not present them as though they are.
What a structure-only session actually looks like
Concretely, a session that stays on the right side of this line tends to follow a recognizable shape. You describe the assignment and your rough thesis in your own words. You ask for two or three possible structures, or a list of sections a piece like this typically needs, and you pick and rearrange rather than accepting the first one wholesale. You describe the specific paragraph or transition you are stuck on and ask for approaches, not finished text, then write the actual sentences yourself once you see a path through. You paste in a paragraph you already wrote and ask what is unclear about it, then rewrite the unclear parts in your own words rather than accepting a supplied replacement. Notice that in every step, the words that end up in your document were composed by you, reacting to prompts and structure the tool suggested — the tool never originated a sentence that survived into the final draft.
This connects directly to how long the writing actually takes
Used this way, AI assistance has a real, measurable effect on your writing schedule, and it shows up in exactly the place you would expect: the time you would otherwise lose to being stuck. A solid outline going in means less time spent staring at a blank section wondering what belongs in it, and a quick plain-language explanation of a concept you were fuzzy on can turn a long confused detour into a two-minute clarification. Both of those translate into a genuinely faster sustained words-per-hour pace when you sit down to draft, because you spend more of that hour actually composing sentences and less of it stalled. That is a legitimate way to plan a faster schedule. Generating the sentences themselves does not speed up your writing in any sense worth measuring — it replaces your writing with something else, and the honest version of your schedule would then need to include the time you will spend rewriting or verifying whatever the tool produced before you could respect it enough to submit it.
A quick self-check before you keep a generated sentence
When you are unsure whether something you produced with AI assistance belongs in your document, ask two questions. Could you have written this sentence yourself, given enough time, using only your own understanding of the material? And can you defend every claim in it, right now, without opening the conversation with the model again? If either answer is no, the sentence is not ready to be yours yet — either rewrite it from your own understanding, or go verify the claim it contains before it goes anywhere near a submission. This is a slower process than accepting generated prose wholesale, and that is exactly the point: the slowness is where the actual thinking, and the actual learning, happens.
Where this leaves genuinely difficult writing
None of this makes hard writing easy, and it is not supposed to. A well-structured outline and an unstuck transition remove friction that was never the valuable part of writing in the first place — nobody's understanding of a subject improves from staring blankly at a cursor. What is left after that friction is removed is still the actual work: deciding what you think, finding the evidence, and building the sentences that carry your reasoning. AI used for structure and stuck points clears the runway. It does not, and should not, do the flying.
If your institution asks you to describe your AI use, this is usually the easiest kind of use to disclose honestly and specifically, because it does not require you to account for how many of your final sentences came from the tool — the answer is none of them, and you can say so plainly, along with what the tool actually did help with, which is a much more comfortable statement to write, and to stand behind, than a vague one.