How Writers Can Maintain Originality When Using AI Tools

Originality is a slippery word, and the AI argument has made it slipperier. It is worth separating three things that usually get argued about as if they were one.

The first is provenance: who or what produced the words. This is the question detection tools, publishers and university policies are asking, and it has a factual answer.

The second is originality of thought: whether the piece contains an idea, an argument or an observation that did not exist in this shape before. A model, by construction, is a very good average of what has already been written. It can recombine, it can summarize and it can produce a competent version of the consensus. What it cannot do is have spent nine years teaching seventh-graders and noticed the one thing about them that nobody writes down.

The third is voice: the texture that lets a reader recognize you across two paragraphs without a byline. Voice is made of rhythm, word choice, what you linger on and what you refuse to explain.

An AI tool threatens the second and third far more than the first. You can be entirely transparent about using a model and still publish something with no thought and no voice in it. That is the failure worth guarding against, because it is the one readers actually feel.

Knowing where a model smooths things over makes editing much faster, because you stop reading for quality and start reading for specific defects. These are the ones that show up in almost every generated draft.

• Averaged sentence rhythm. Human paragraphs vary wildly, with a nine-word sentence landing after a forty-word one. Model output settles into a middle length and stays there, which reads as competent and forgettable.
• Symmetry where life is asymmetric. Three examples per point, two clauses per sentence, a summary that restates the opening. Real arguments are lopsided because some parts matter more.
• Hedged conviction. A model will tell you a thing is often considered important. A writer tells you it matters and then says why they think so.
• Vocabulary that belongs to nobody. Words like leverage, robust, seamless and delve are not wrong, they are simply the words with no fingerprints on them.
• Missing specifics. The draft names a category where a writer would name a thing: a book instead of a battered library copy of Dubliners with someone else's pencil marks in it.

None of this makes a draft unusable. It makes a draft a draft. The mistake is treating fluent output as finished output, because fluency is exactly the thing a language model is optimized to produce, and it is the thing that most resembles being done.

The single highest-leverage habit is also the least glamorous: write the structure yourself, before you ask for a single sentence. Structure is where the thinking lives. If you hand the model a topic and let it decide what the piece argues and in what order, you have outsourced the part that makes the work yours and kept the part that is mostly typing.

An outline you own does not need to be elegant. It needs four things.

• The claim you are actually making, written in one sentence you would defend out loud.
• The order of the argument, and why it is that order rather than another one.
• The examples, anecdotes and sources you already own, marked where they go.
• The places you genuinely do not know yet, marked as gaps to research rather than gaps to generate.

With that in hand, drafting with a tool stops being generation and becomes expansion. You are asking it to render a decision you already made, which is a much narrower job and a much better one for a model to do. For long-form work, this is also how you keep a book coherent across forty chapters, since the argument is fixed somewhere outside the model's memory of the last three paragraphs.

There is a meaningful difference between editing a generated draft and rewriting it. Editing keeps the model's sentences and tidies them, which preserves the rhythm that made the draft feel anonymous in the first place. Rewriting keeps the draft's information and replaces its sentences with yours.

The practical version: read a generated paragraph, look away from the screen, and write the paragraph you would have written if the draft did not exist. Then check the draft to see whether you missed anything. This is slower per paragraph and dramatically faster overall, because you stop producing prose that sounds fine and needs three more passes to stop sounding fine.

Five specific moves that do most of the work:

• Delete the first sentence of every generated paragraph. It is usually setup the reader does not need.
• Replace every abstract noun with the concrete thing it stands for.
• Break one long balanced sentence into a long one and a short one.
• Put back the opinion the model hedged away, in your own words.
• Add the detail only you could have supplied: the client who said the annoying thing, the draft you abandoned, the number you looked up yourself.

A useful rule of thumb from working writers: if you have not changed at least half the words in a generated paragraph, you probably have not thought about it yet.

Describing your voice to a model rarely works. Asking for warm, conversational and a little dry gets you the internet's average of those adjectives. Showing the model your actual prose works considerably better, because it is matching a pattern rather than interpreting a label.

Keep a plain text file of your own writing and paste it in at the start of every session.

• Use 500 to 1000 words of your own finished prose, not a description of your style.
• Pick a passage that is typical rather than your best writing, so the model matches your baseline instead of your peak.
• Keep separate samples for separate registers: your newsletter voice is not your book voice.
• Refresh the sample once a year. Your writing moves and a stale sample pulls you backwards.

Expect this to get you part of the way rather than all of it. A style sample reliably fixes register, sentence length and vocabulary. It does not supply judgement, and it will happily imitate your cadence while saying something you do not believe. That remains your job.

Reading aloud is an old copyediting trick and it happens to be the best available detector of machine-flattened prose. Your eye forgives an even rhythm. Your mouth does not. Sentences that all breathe at the same interval become obvious within a paragraph, and so do the transitions that exist only because the model likes transitions.

Read a full section, not a sentence, and mark three things: where you ran out of breath, where you got bored, and where you found yourself adding a word that is not on the page. That last one is the most valuable. The word your mouth supplies is almost always the word your voice wanted, and it is almost always more specific than what the model wrote.

Detection belongs at the end of the process, not the middle, and it answers a narrower question than most people assume. A detector estimates how statistically predictable your text is. That correlates with machine authorship, but it is a correlation and not a verdict, and plain, careful, well-structured human prose is flagged often enough that no writer should treat a score as a judgement on their work.

Used properly, it is still genuinely useful, because the passages it flags are usually the passages you rewrote least. Before a piece goes to an editor, a client or a publisher, it is worth running the finished draft through an AI detector and reading the flagged sections back to yourself, not to chase a lower number but to find the paragraphs where you stopped paying attention. We ship our own AI content detector for the same purpose, and the advice is identical whichever you use.

Two cautions. Do not edit to defeat a score: rewriting to look unpredictable produces worse prose than rewriting to sound like you, and the two are not the same exercise. And if you are writing under a policy that asks about AI use, answer that question with disclosure rather than with a detector result. Provenance is something you know and a tool only guesses at.

The writers who keep their originality while using AI tools are not the ones with better prompts. They are the ones who never gave away the outline, who rewrite instead of accepting, who keep a sample of their own prose within reach, and who read the thing out loud before anyone else sees it. The tool drafts. The judgement, the argument and the specifics stay with the person whose name goes on the cover.

That is a workflow, not a compromise, and it is considerably faster than writing every sentence from nothing. The relief of the blank page disappearing is real. You just have to spend some of the time it saves on the parts a model was never going to do for you.