Keep Your Voice When Using AI to Write

Keep your voice when using AI to write by treating the model as a drafting assistant, not a co-author. You write the outline, you decide which sections AI may draft, and you rewrite every sentence with details only you can supply. Below is the exact loop, plus how to tell whether the finished draft still sounds like you.

Key takeaways

• Outline in your own words before any AI tool touches the draft. The outline is the part that carries your argument.
• Let AI draft only sections you can already explain out loud, without notes.
• Edit for specificity you alone can supply: names, places, figures, moments from your own reading or fieldwork.
• Check the finished draft by reading it aloud and by scoring it sentence by sentence, not just as a whole document.
• Stop using AI on a piece once you can no longer defend every claim without checking what the model said.

Why does AI drafting flatten your voice in the first place?

Language models predict the next likely word. That produces prose that is competent, smooth and statistically average. It gravitates to the middle of the distribution: general verbs, abstract nouns, balanced sentences of similar length.

Your voice lives at the edges of that distribution. It is the odd concrete detail, the sentence you cut short because the point was already made, the example nobody else in the seminar would have chosen. A model cannot reach for those, because they are not in its training data. They are in your notes.

This is why the problem is not that AI writes badly. It is that AI writes plausibly, and plausible is the enemy of memorable. If you accept a full draft and lightly polish it, you have sanded off the only parts that were yours.

Our own measurement makes the mechanical side visible. On 2026-09-19 we scored two corpora with the DeAIze detector (local ensemble, lite tier) at a 30% flag threshold. A set of 46 hand-written DeAIze guides had a mean AI score of 27% and a median of 25%. Twelve unedited model passages, generated from fixed prompts and never touched, had a mean of 35% and a median of 34%. The two ranges overlap. That overlap is the whole story: detectors are estimating probability, not detecting authorship, and a human draft can sit in the same band as raw model output.

The co-writing workflow, step by step

Step 1: Outline in your own words, before any AI touches it

Open a blank document. Write the argument as a list of claims, not topics. Not “paragraph about the counterargument” but “the counterargument fails because the sample was drawn from one region.” One line per beat.

Do this with the model closed. If you open ChatGPT first, you will unconsciously negotiate with its suggestions and end up with its structure. The outline is the load-bearing wall of the piece.

A useful test: can you say the outline out loud in under two minutes? If not, you do not yet know what you are arguing, and no amount of drafting will fix that.

Step 2: Mark which sections AI may draft

Go through the outline and label each beat. I use three labels:

Label	Meaning	Who drafts it
Mine	The argument, the analysis, the conclusion	You, alone
Assisted	Background, definitions, a first pass at a transition	AI drafts, you rewrite
Mechanical	Formatting, a table shell, a citation list	AI drafts, you verify

The rule that keeps this honest: only let AI draft a section you can already explain out loud. If you cannot summarise the point in your own words at the kitchen table, you are not editing the model’s draft, you are inheriting it. That is how authorship quietly transfers.

Step 3: Prompt for the beat, not the piece

Give the model one beat at a time, with the surrounding outline as context. Ask for two or three variants. You are not looking for the right sentence; you are looking for a shape you can push against.

A prompt that works: “Here is my outline. Draft only section three, in about 150 words, plain register, no headings. Do not add claims that are not in the outline.”

That last instruction matters. Models pad. They add a clause about the broader significance of the topic, and that clause is usually the first thing a reader skips.

If you want to see which lines in your own draft carry the machine signal, run it through the AI detector first. It scores the document and highlights each sentence, so you can fix the flagged lines instead of rewriting everything.

Edit for specificity only you can supply

This is where the piece becomes yours. Go sentence by sentence through every AI-drafted passage and ask one question: could anyone have written this sentence?

If yes, replace something in it with something only you have. Not decoration. Evidence.

• Names. The author you actually read, not “researchers have argued.”
• Numbers. Your own counts, dates, measurements, prices, page references.
• Examples. The specific case from your placement, your lab, your part-time job, your reading group.
• Friction. The thing that did not work, the objection you could not answer, the source that contradicted you.

Here is an illustration. Machine-drafted: “Many students report feeling anxious about using AI tools in their coursework, which can affect their academic performance.” Rewritten with specificity: “In my seminar, three of the eight people who admitted using ChatGPT for essays said they rewrote the whole thing anyway, because they could not defend a paragraph they had not written.”

The second version is not more elegant. It is more true, and it is checkable. That is what voice actually is at this stage of a career: a record of what you noticed.

How do you check whether the final draft still sounds like you?

Three passes, cheapest first.

1. Read it aloud. You will hear the flat stretches. They are the ones where your mouth keeps going but your attention drifts. Mark them.
2. Cover the byline test. Give two paragraphs to a friend who knows your writing and ask which one sounds like you. If they hesitate, you have your answer.
3. Score it at sentence level. Paste the draft into the AI detector and look at the highlights, not the headline number. A whole-document score tells you very little. Which sentences carry the signal tells you where the generic prose is hiding. If you want the background on what those scores mean, our guide to how AI detection works walks through the signals.

One honest caveat. A high score is not proof you used AI, and a low score is not proof you did not. Detector output is a probability estimate. Treat it as a map of where your prose is bland, and ignore it as a verdict on your integrity.

When should you stop using AI on a piece entirely?

Stop the moment any of these becomes true.

• You cannot explain a paragraph without rereading what the model wrote.
• You have started accepting sentences because they sound fine, not because they are right.
• The piece contains a claim you cannot trace to a source you have actually read.
• You are past the point where you could defend the argument in a viva, a hearing or a client call.
• The deadline is close enough that you will not have time to edit properly. An unedited AI draft is worse than a shorter piece you wrote yourself.

There is also a structural reason to stop early on some pieces. If your course or employer has a disclosure rule, read it before you draft, not after. Our breakdown of what universities actually allow is a reasonable starting point, and the rule of thumb is simple: the ideas, data and conclusions have to be yours. AI can help you phrase a transition. It cannot supply your position.

The fear underneath all of this is worth naming plainly. You are worried that leaning on the model will make you sound like everyone else, and that the moment a detector or an editor looks closely, the work will read as borrowed. That fear is reasonable. It is also manageable, because the fix is not to hide the AI use. The fix is to put so much of yourself into the draft that the question stops being interesting.

What a finished co-written paragraph looks like

Compare the two versions below. Both make the same point. Only one could have been written by a specific person on a specific day.

	Version A (model-drafted, lightly edited)	Version B (co-written, specificity added)
Opening	“AI tools are increasingly common in academic writing.”	“By the second term, half my cohort had a ChatGPT tab open during seminars.”
Evidence	“Studies suggest mixed results.”	“The two people who showed me their drafts had both deleted the model’s conclusion.”
Ending	“Students must use these tools responsibly.”	“The tool was never the problem. Not knowing what I thought was.”

Version B took longer. It also survives a detector flag, a supervisor’s question and a reread six months later, because every sentence has a source: you.

If you want the fuller version of the editing pass, the humanize AI text pillar collects the related guides, and why AI-assisted drafts read flat goes deeper on the mechanics of the fixes.

A short checklist before you submit

• The outline was written with the model closed.
• Every AI-drafted section is one you could explain out loud first.
• At least one specific detail in every AI-assisted paragraph: a name, a number, a date, an example from your own work.
• The conclusion was written by you, from scratch, after the rest was finished.
• You read the whole thing aloud once.
• You checked the sentence-level highlights rather than trusting a single document score.
• You can point to every claim and say where it came from.

Seven items. If you cannot tick the last one, the rest do not matter much.

Ready to see which sentences read machine-made?

You do not need a subscription to find out. Paste your draft into the AI detector and read the sentence-level highlights. Then, if the flagged lines are ones you wrote yourself, run them through the draft editor to see what changes and what survives. If you are working across a longer document, the pricing page shows what each word allowance covers, and you can start with the free tier without a card at register.

The workflow above costs nothing and works with any tool, including none. Use it either way.