How to edit AI-written text so it stops reading like AI

Detectors that claim to tell you whether a text was machine-written are unreliable in both directions: they flag careful human prose and wave through edited model output. Readers are a different matter. People who read a lot of text online have learned the tells without being able to name them, and the reaction is not "this was written by a model" so much as "I stopped trusting this". The fix is an editing pass, not a detector — and most of it is mechanical enough to run in a browser tab.

The vocabulary layer

Start with the phrases. Models are trained toward the safe middle of everything, which means they reach for the same connectives and the same reassurances over and over. The recurring set is small and very recognisable: it's important to note that, let's delve into, in today's fast-paced world, a testament to, navigate the complexities of, plays a crucial role in, rich tapestry, at its core, plus the conversational scaffolding a chat interface adds on top — certainly, great question, I hope this helps.

Almost none of these carry meaning. Delete it's important to note that and the sentence after it says the same thing with more authority. The AI phrase remover matches that catalogue and strips it, then repairs the capitalisation and spacing the deletion leaves behind, which is the part that makes doing this by hand tedious. Run it first, before you edit anything else, because roughly a tenth of a typical draft disappears and the shape of the remaining text changes.

Clichés are the neighbouring problem and need a different treatment. Game changer, low-hanging fruit, the elephant in the room: these are not machine-specific, they are just dead. The cliché detector marks them rather than deleting them, because a cliché sometimes is the right call in a sentence and only you can tell. Two or three in a long article is style. Nine is a draft nobody thought about.

The rhythm layer

Vocabulary is the easy half. What actually gives a draft away is that every paragraph is the same length, every section has three bullets, and every list has three items. Models produce triads compulsively — three adjectives, three examples, three clauses — and once you notice it in a piece you cannot stop noticing it.

Other structural tells worth hunting by eye:

• Negative parallelism. "It's not just X — it's Y." Once is fine. Four times in a page is a verbal tic.
• The -ing summary clause. Sentences that end with "…, highlighting the importance of collaboration" or "…, underscoring its role in modern workflows." These add no information; they restate the sentence in a more abstract register. Cut the clause and the sentence survives intact.
• Symmetrical openings and closings. An intro that announces what the piece will cover and a conclusion that recaps what it covered, with nothing in either that the middle does not say better.
• Uniform paragraph length. Human writing lurches. A three-word paragraph after a long one is a rhythm change no model produces spontaneously.

None of these is detectable by string matching, which is why this part stays manual. What you can automate is the diagnosis: run the draft through the readability score and look at the sentence-length distribution rather than the grade number. A flat distribution — every sentence between eighteen and twenty-four words — is the signature. Human prose has a much wider spread.

Passive voice, but for the right reason

Model output leans passive because passive constructions avoid committing to who did something. "Mistakes were made in the deployment process" is a sentence about nothing. The passive voice detector lists the passive sentences and, usefully, shows the agent when one is present — so you can see at a glance which passives hide the actor and which are legitimate.

Do not convert all of them. Passive is correct when the actor is unknown, irrelevant, or deliberately unnamed, and English scientific writing uses it as a convention. The rule of thumb: if you can add "by whom?" to a passive sentence and the answer matters to the reader, rewrite it. Otherwise leave it and move on.

The invisible layer

Text copied out of a chat interface carries characters you cannot see. Non-breaking spaces instead of ordinary ones, zero-width joiners left over from emoji sequences, soft hyphens, narrow no-break spaces around punctuation, and occasionally a Cyrillic character sitting inside a Latin word after a round trip through some intermediate tool. These survive a paste into a CMS and then break search, break diffing, and produce the mystery where two apparently identical strings do not compare equal.

The invisible character detector shows you every one of them with its code point and offset, so you can decide what to remove rather than blindly stripping everything — a non-breaking space is often deliberate. Do this last, after all the rewriting, because every editing pass through another tool can add its own.

What not to do

Two overcorrections are worse than the original problem.

The first is purging the em dash. It became a supposed AI marker, and people now hunt it in their own writing — but the dash is a normal English punctuation mark and replacing every one with a comma produces prose that reads like a form. What makes dash usage look machine-generated is uniformity: one per paragraph, always in the same rhetorical position. Vary the punctuation, don't ban it.

The second is running the text through a "humaniser" that swaps words for synonyms. That produces something worse than the original: text with the same flat structure and now-wrong word choices, which reads as machine-written and careless. Structure is where the tell lives. Synonyms are decoration.

The pass, in order

Remove the stock phrases. Mark and thin the clichés. Read the sentence-length distribution and deliberately break it — merge two sentences somewhere, split another, leave one paragraph very short. Fix the passives that hide an actor. Cut every trailing -ing clause that restates the sentence. Then sweep the invisible characters and copy it out.

Fifteen minutes on a page, and the result is not "undetectable" — it is simply better written, which was the actual goal. Every one of these tools runs locally in your browser, so an unpublished draft stays on your machine while you work on it.