First Drafts That Do Not Look It

AI-assisted writing is identifiable, and not for the reasons commonly given. Banned-word lists miss the actual signals.

What actually gives it away

Structural uniformity. Every section the same length, every paragraph the same shape, three examples per point. Human writing is lumpy — one idea gets four paragraphs and the next gets a sentence.

Symmetrical balance. Every claim immediately hedged, every advantage paired with a disadvantage, every position given equal weight. This reads as fairness and is actually an inability to have a view.

Empty transitions. Sentences whose only function is moving between paragraphs.

Restating the question. Opening by rephrasing what was asked.

Summary paragraphs that add nothing. A closing section restating what was just said.

Generality where specificity was needed. The classic tell. Real writing contains details only that writer would have: a number from your own data, a thing that happened, a case you handled. Generated text contains the category where the example should be.

Relentless competence with no texture. No aside, no admission of uncertainty, no sentence that took a risk.

What does not give it away

Particular words. The lists circulating are unreliable. Perfectly ordinary writers use these words, and avoiding them produces stilted prose.

Em dashes. A punctuation mark used by writers for centuries.

Good grammar.

Optimising against a word list makes writing worse without making it less detectable.

A workflow that produces usable drafts

Do the thinking first. Decide what you are arguing before generating anything. A draft produced from a vague prompt is generically shaped because the input was.

Supply the specifics. Your data, your examples, your constraints, what you actually think. The model can structure and phrase; it cannot supply the material only you have.

Ask for structure before prose. An outline you can argue with is more useful than paragraphs you have to unpick.

Generate in pieces. Section by section, with your input between. A whole document in one pass has the uniform shape.

Then rewrite the opening and the closing yourself. These carry the most voice and they are where generated text is most formulaic.

Break the symmetry deliberately. Cut the balancing clause where you actually have a view. Let one section run long. Remove the closing summary.

Add one thing only you could have written. A specific number, a specific incident, a specific objection you have encountered. This does more than every other step combined.

Editing rather than generating

The stronger use, and the underused one.

Write badly and fast yourself, then ask for a critique: what is unclear, what is unsupported, where does the argument break, what would a hostile reader attack.

This keeps your structure and your voice while getting the benefit of a second read. The output is your writing, improved, rather than generated text you have to disguise.

For anything where voice matters — anything with your name on it — this is the better approach by a wide margin.

Disclosure

Norms differ by context and are still forming.

Academic work: follow the institution's rules, which usually exist now and usually require declaration.

Professional work: the question is generally whether the output is correct and whether you take responsibility for it, not how it was produced.

Journalism and publishing: many outlets have policies. Check.

Anything with a byline: the byline is a claim of authorship. If you did not do the thinking, that claim is weaker than it looks.

The consistent principle: you are responsible for what you put your name to. Every fact, every claim, every citation. "The model wrote it" is not a defence anywhere, and the person who discovers the fabricated reference will not be sympathetic.

The output that is genuinely fine

Routine correspondence. Meeting notes. Documentation. Summaries. Boilerplate. Things where nobody expects voice and the value is entirely in accuracy and clarity.

Using a model for these is straightforwardly good. The care described above applies to writing where the voice is part of the point — which is a much smaller share of what most people write than they assume.

Instructions that reduce the generic shape

Prompting cannot eliminate the tendencies described above, and it can reduce them noticeably.

"Do not include an introduction or a conclusion." Removes two of the most formulaic sections outright.

"Vary the paragraph lengths." Directly targets the structural uniformity.

"Do not hedge. Where the evidence supports a position, state it." Cuts the symmetrical balance.

"Use my examples only. Do not invent examples." Prevents the generic illustration.

"No transitional sentences."

"Write as if for someone who already knows the basics." Removes the explanatory padding that pads out most generated prose.

Give it a sample of your own writing and ask it to match the register. This works better than adjectives about tone.

Then still rewrite the opening yourself. No instruction fixes the first paragraph, and the first paragraph is what the reader judges.