What "sounds like AI" actually means
Readers are not detecting a model. They are noticing an absence: no commitment, no specifics, no variation in rhythm. Language models were trained to produce text that is broadly acceptable, so their default output sits at the centre of every distribution — medium sentences, hedged claims, balanced structure. Human writing done in a hurry, or written to fill a required length, drifts to exactly the same place. That is why the effect appears in drafts nobody generated.
The seven signals below are the ones the writing tool on this site measures. Read them as a diagnostic list for a draft, not as a test of a person.
1. Uniform sentence length
The strongest single signal. When almost every sentence lands between fifteen and twenty-five words, prose develops a metronomic pulse that readers register as artificial. Human writing under real editorial pressure varies far more, because some ideas need a clause and others need a paragraph.
Revision: read a paragraph aloud and split the longest sentence in two. Then cut one sentence to under six words. A short sentence after a long one is the cheapest available emphasis, and it immediately breaks the pulse.
2. Repeated vocabulary in a narrow band
Not overall repetition — keyword repetition is normal and useful — but the same abstract nouns cycling through every paragraph: approach, framework, solution, process, aspect, factor. These words are placeholders where a concrete noun belongs.
Revision: for each abstract noun, ask what it refers to. "The onboarding process" becomes "the four forms new hires fill in on day one". Specificity removes the repetition automatically.
3. Connector density
Furthermore, moreover, additionally, in addition, however, therefore. One or two per page are invisible; one per paragraph turns an argument into a list wearing a suit. Heavy connector use usually signals that the underlying order is arbitrary, so each paragraph needs a word to explain why it follows the last.
Revision: delete every connector, then reread. Wherever the sequence still makes sense, leave it deleted. Where it does not, the fix is to reorder the paragraphs rather than to reinstate the word.
4. Template phrasing
"It is important to note that", "plays a crucial role in", "in today's rapidly evolving landscape", "when it comes to". These carry no information; they occupy the position where a claim should be.
Revision: delete the phrase and keep the sentence. "It is important to note that costs rose" is "costs rose". If the sentence collapses to nothing, the paragraph was decoration.
5. Repeated sentence openings
Three consecutive sentences beginning with "This", or a run of paragraphs each opening with the subject of the section, produce a flatness readers feel before they can name it.
Revision: scan the first two words of every sentence in a section. Vary at least half of them — open occasionally with a time marker, a dependent clause, or a direct question.
6. Regular punctuation rhythm
Prose consisting only of full stops and commas reads more evenly than prose that uses the full range. Colons, dashes, semicolons and the occasional question mark are pacing instruments.
Revision: introduce one colon where a sentence announces a list or a consequence, and one dash for an aside. Do not add more than that; punctuation variety is seasoning, not substance.
7. Even paragraph lengths
Paragraphs of near-identical length suggest the text was produced to fill a shape rather than to make a point. Real arguments are lumpy: a claim may take six sentences to establish and one to qualify.
Revision: find the paragraph that carries the most important idea and let it run longer. Isolate one crucial sentence as its own paragraph.
The revision that beats all seven
Add something only you could have written. A number you measured, a date, a name, a decision you regret, a counter-example that weakens your own case. Specific detail cannot be produced by averaging a corpus, and it simultaneously fixes vocabulary repetition, template phrasing, and paragraph uniformity, because concrete material has its own shape.
The corollary matters too: do not add randomness to defeat a score. Inserting typos, unusual synonyms, or arbitrarily fragmented sentences makes writing worse while moving the number. Any tool that promises to "humanise" text by shuffling it is optimising the metric and damaging the document.
If you were accused rather than curious
People arrive at this page because someone ran their work through a detector and it produced a high number. That number is not evidence. These signals are statistical properties of a sample, and they are elevated in second-language writing, in technical documentation, in text written to a rigid template, and in careful formal prose. No public checker — including the one on this site — can determine who wrote something.
What does help is process evidence: draft history, document version history, notes, and the ability to talk through your reasoning. Keep those; they are far more persuasive than a rebuttal score. The reasoning behind this limit is set out in what an AI writing checker can actually tell you.
Using the tool on this site
The writing panel returns a reference score plus the individual signal readings, so you can see which property is elevated rather than a single verdict. Analysis runs in your browser; the text is not uploaded in the default path, and a server model is used only if the deployment configured one and you tick the per-request consent box. Treat a high reading as a prompt to reread the draft, and treat a low reading as no evidence of anything at all.