How to make AI text sound human (what actually changes it)

The patterns that give it away

**Uniform rhythm.** Generated prose tends to hold a steady 15–25 words per sentence. Human writing lurches — a 40-word sentence, then a four-word one. That variance is the strongest single signal, and it is the easiest to fix.

**Hedging before every claim.** "It is worth noting that", "generally speaking", "in many cases". Sampling from high-probability tokens favours the safe qualifier, so hedges accumulate. Most can be deleted with no loss of meaning.

**The rule of three, endlessly.** Three adjectives, three examples, three clauses. Once you see it you cannot unsee it. Break some into two and some into five.

**Summary sentences.** A paragraph makes its point, then a final sentence restates the point. Cut those; they carry no information.

**Recycled abstractions.** "Landscape", "realm", "delve", "tapestry", "testament", "crucial", "leverage". Not wrong individually, but the density is unusual.

**Perfect punctuation.** Em dashes placed correctly every time, curly quotes throughout, no comma splices. Real writing has texture, and some of that texture is small imperfection.

Edits that actually change how it reads

Read it aloud. Anything you stumble over is a sentence you would not have written.

Force length variance deliberately: find your longest paragraph and make one sentence in it under six words.

Delete every hedge, then reinstate only the ones carrying real uncertainty. Usually about one in five.

Replace abstract nouns with the specific thing. "Improved outcomes" becomes "cut the refund rate". Specificity is the hardest thing to fake and the fastest to convince.

Add something only you know — a number from your own work, an objection you actually heard. Generated text cannot supply this, and one sentence of it does more than a page of rephrasing.

What this does and does not affect

**Character-level artifacts are separate.** Zero-width characters, non-standard spaces, smart punctuation and pasted HTML class names have nothing to do with style. They survive every rewrite, because rewriting changes words and these are not words. They have to be removed as bytes, which is what our checker does.

**Statistical classifiers respond to style.** Tools like GPTZero and Turnitin score how predictable your word choices are, so genuine variance in structure and vocabulary moves that score. They remain unreliable in both directions, and a lower score is not proof of anything — that unreliability is the point, not a feature to depend on.

**Watermarks are a different mechanism.** A distributional watermark is a keyed bias over token choice. As a general property, substantive paraphrase degrades such marks while retyping does not, since retyping reproduces the same words. But the honest position is that nobody outside the provider can verify a mark either way: no public detector exists, and Anthropic's stated scope covers models launched on or after 2 August 2026, which no shipped model currently meets. We do not claim to remove Anthropic's watermark, and you should be sceptical of anyone who does.

The thing most rewriting advice gets wrong

Swapping words for synonyms makes text worse and no more human. It preserves the structure that was the actual signal while degrading the meaning, which is why heavily "humanised" text often reads stranger than the original.

Structure is the signal. Change the rhythm and the specificity and the text reads as written by a person, because at that point a person has genuinely rewritten it.

Questions

Do AI humanizer tools work?

They lower classifier scores, which is not the same as reading as human. Most work by synonym substitution and clause reordering, leaving the structural patterns intact — and they can introduce awkward phrasing that makes the writing worse.

Does rewriting remove the Claude watermark?

Nobody can verify this either way, because there is no public detector. As a general property of distributional watermarks, substantive paraphrase degrades the signal and retyping does not. Treat any product that promises watermark removal as unverifiable, including on this point.

Will editing for style remove invisible characters?

No. Those are bytes, not style, and they survive any amount of rewriting. They need to be stripped separately.

Is the em dash really an AI tell?

On its own, no — Word and Google Docs both produce em dashes automatically from text you typed. It is only suggestive alongside the structural patterns.