AI sounding text is a handful of constructions, distributed perfectly evenly.

People who say a page sounds like AI are not describing a statistical property. They are recognising a set of moves: the opening that addresses the whole internet, the contrast that says it is not a tool but a way of working, the promise that the key is something, the second person that could apply to anyone.

Each of those moves exists in good writing too. What makes text sound generated is the frequency and the evenness: all four appear in the same paragraph, in the same order, at the same weight, in page after page.

The reaction is stronger than the evidence deserves, which is worth knowing before you argue with it. A reader who suspects the text was generated reads it differently, looking for reasons to leave rather than reasons to stay.

So the goal is not to hide that a model was involved. It is to remove the moves that add nothing, so the page can be judged on what it says.

Four families cover most of what people notice. All are ordinary rhetorical devices. The problem is that they arrive together and unchosen.

Models have stylistic habits beyond vocabulary. A heavy reliance on the em dash for asides, colons that introduce a three-item list, and a fondness for the balanced semicolon are all common.

None of these is wrong, and each is a marker of competent writing by some standards. The tell is striking them evenly: an em dash in every other paragraph and never a parenthesis, three-item lists everywhere and never a two-item one.

The tricolon is the clearest example. Everything arrives in threes, because three is the shape that sounds complete. Use two when there are two things and five when there are five, and the page immediately reads like it was written by someone counting.

Check your quotation marks and spacing too. A draft that uses curly quotes inconsistently, or mixes one and two spaces after a full stop, is telling you it was assembled from more than one source.

Generated prose gives every point the same weight. The central claim and the incidental clarification occupy the same paragraph length, use the same sentence structure and receive the same amount of hedging.

A reader reads that as flatness, because emphasis is how they know where to pay attention. If nothing is emphasised, everything is equally skippable, and the reader skims or leaves.

Emphasis is a decision: one sentence per section gets a short, unqualified statement of the point. The rest can be qualified, extended and complicated.

The test is whether you can point at the sentence in a section that a reader must not miss. If you cannot, the section has no point and the evenness is hiding that.

People are good at recognising constructions they have read a thousand times and bad at estimating whether a specific page came from a model. Detectors are the reverse: they output a number and have no idea what it means to a reader.

The disagreements are predictable. A human-written page in a formulaic genre, such as a how-to list or a press release, triggers every construction and reads as generated. A lightly edited model draft that follows an unusual structure reads as human.

This is why the phrase sounds-like-AI is best treated as an editorial note rather than an accusation. It means the writing is doing things the reader has learned to distrust, which is actionable whatever produced it.

The corrections are also the same as they would be for a bored human writer: cut the formula, add a specific, vary the shape. Nobody has to establish authorship to improve the page.

Formal register is not a tell. Technical vocabulary is not a tell. Full sentences, correct grammar and consistent terminology are not tells, though detectors often score them as suspicious because the text is predictable.

Non-native English writing is the most common false signal. Careful, slightly formal prose from a competent second-language writer looks exactly like the output of a model trained to be careful and formal, and it is the single largest source of bad accusations.

Translated text has the same problem. A good translation preserves the source's structure, which often differs from idiomatic English rhythm, and the result reads as machine-written to anyone scanning for variety.

Before you tell someone their writing sounds generated, check whether you are describing the writing or the person. The first is useful. The second is not, and it is usually wrong.

Take a paragraph with three of the constructions in it. Delete the opening formula entirely and start with the first fact. Nine times out of ten the paragraph is better and shorter.

Then convert the contrast formula into a plain statement. Instead of this is not just a checklist, it is a process, write the checklist has four steps and the process decides their order. Same idea, no formula, and now the reader knows something.

Then remove one instance of universal address. Whether you are a solo founder or a large team becomes the smaller case is where it breaks, which is specific, arguable and useful.

Finally, cut one sentence's worth of hedging and read the paragraph aloud. If it now sounds like one person making a point to another, the rewrite worked.

Some of these devices are fine, and removing all of them produces a different failure: a page so determined to avoid a formula that it cannot make a clean argument.

Keep a contrast formula when the contrast is the point, once per page. Keep a three-item list when there are exactly three things. Keep second person when you are genuinely addressing one kind of reader.

The rule is one per page, chosen deliberately, and preferably not in the opening. A construction that appears once reads as style. The same construction in the first and last paragraph reads as a template.

Delete the rest and see whether anything is lost. In most drafts nothing is, which is the clearest evidence that the constructions were carrying no meaning.

The most reliable test is a person who does not know what the page is for and has not seen your outline. Give them the first two paragraphs and ask what they expect from the rest.

If they guess correctly in a way that makes the rest redundant, the opening was formulaic and told them nothing. If they ask a question the page answers three sections later, the structure is wrong but the content has something.

If they say it sounds like AI, ask which sentence. The answer is almost always one of four constructions, and knowing which one is worth more than any percentage score.

Run this on pages you have not written yourself. The author is the worst judge of their own cadence, because they can still hear the intention behind the words.

One: does the first sentence state a fact or announce the article? Two: is there a not-just-X-but-Y in the first hundred words? Three: how many sentences in the draft are about the same length?

Four: is there a single sentence a reader must not miss, and is it short? Five: which constructions appear more than twice? Six: would a reader who does not know you say this sounds like one person or like a category?

Six questions, five minutes, on a page you have read four times. That is the whole method, and it catches most of what makes a reader stop trusting a page.

The scanner on this page does the counting so you can spend the time on the decisions.