How to avoid AI detection
Detectors measure four properties of text. Change those and the score moves. Change anything else and it does not, which is why most advice on this topic fails.
Almost everything written about this is guesswork. The properties detectors read are known and countable, so here they are, in order of how much each one moves a score.
1. Sentence length variation
The strongest single signal, and the one almost nobody works on.
Human academic writing in our corpus varies its sentence length with a standard deviation of 8.7 words. Generated text sits at 6.4. Generated prose is not too long or too short, it is too even.
What to do. Find a run of three sentences of similar length. Split one and merge another pair. A four-word sentence next to a thirty-word one does more for this number than an hour of word substitution.
2. Connective openings
Sentences that begin "Moreover", "Furthermore", "Additionally", "Overall", "Consequently".
These open 2.4% of human sentences in our corpus and 5.8% of generated ones. More than double, and they are visible to a reader as well as to a classifier.
What to do. Delete most of them. If the logical relationship is real, the sentence order already carries it. If it is not, the connective was covering for that.
3. Specific detail
Generated prose generalises. It reaches for the category where a person names the thing. "Various approaches have been proposed" instead of naming two. "Significant improvements" instead of the number.
What to do. Put the specifics back. The method, the figure, the actual example. Specificity is the strongest human signal there is and it is the first thing lost while tightening a draft.
4. Paragraph shape
Generated paragraphs run to a template: topic sentence, two or three supporting sentences, a closing sentence restating the opening. Repeated down a page, the shape itself is a signal.
What to do. Let one paragraph be two sentences. Let another run long. Start one in the middle of the argument without announcing it.
What does not work
Synonym substitution. Changes vocabulary, leaves the rhythm untouched, and the rhythm is where most of the signal lives. It also degrades the prose. See do paraphrasing tools beat AI detection.
Prompt tricks. Instructions like "write in a human style" move the structural properties far less than people expect, because the model's sampling behaviour is what produces the evenness. See ChatGPT prompts to avoid AI detection.
Deliberate typos. Detectors do not score correctness. This costs you a reader and buys nothing.
Unusual punctuation and invisible characters. Not what is being measured, and trivially stripped.
Another polish pass. This one moves the score the wrong way. Editing smooths, smoothing removes variation, so a carefully polished draft scores higher than a rough one. See why good writing gets flagged more.
Measure it rather than guessing
Every change above is local, so a single document percentage hides whether it worked. Score per paragraph, fix the two or three that sit highest, and score again.
Ours is free and unlimited with no account, takes .pdf, .docx and .tex, has no word cap so you can re-run as many times as you like, and never stores your text. Our wrong-flag rate on genuine human writing is 0.4%, which is 99.6% specificity, against Turnitin's 4%.