why does my writing get flagged as ai
detectors don't care if your facts are right. they care about shape. your writing flags when its patterns drift toward the statistical centre — where models write from.
you ran your last linkedin post through a detector and it came back at 74% ai. you wrote it yourself, mostly. so what happened.
the detector isn't lying. it's seeing something real. the question is what exactly it's seeing, and whether you can fix it.
the three collapse patterns
language models optimize for probability, not specificity. they produce what reads well in the centre of a distribution — clean sentences, safe transitions, category-level nouns. over three to five ai-assisted drafts, most writers unconsciously start matching that centre.
we saw this across our voice profiles. three patterns, every time:
sentence-length variation flattens. transitions become generic. concrete nouns get replaced by category nouns. these aren't style preferences — they're mechanical artefacts of how models generate text. they account for roughly 70% of flagged outputs.
sentence length is your fingerprint
your natural sentence-length distribution is one of the most reliable signals of how you think. if you tend toward 8-word punches followed by 35-word explanations, that's not a quirk — that's your rhythm. it signals how you move through ideas.
ai drafts collapse that. sentences cluster between 15 and 22 words because variation is expensive in probability space. the model defaults to what reads cleanly, and clean reads mean average-length sentences with standard structure.
the writing still makes sense. it just stops sounding like you.
90 seconds. paste your last three posts into a word count tool. look at sentence length per post. if the range has narrowed and the average has crept toward 18 words, you've already drifted. — hyv voice audit methodology
your transitions disappear first
every writer with a recognisable voice has phrases they return to. "which means," "here's where it gets complicated," "the part nobody talks about." these carry your thinking shape. readers feel them even when they can't name them.
ai drafts replace them with safe defaults: "however," "additionally," "in conclusion." the writing stays coherent. it stops being yours.
this one is fixable without touching anything else. find your three most-used transitions before you open the ai. check whether they survived the draft. if they disappeared, put them back.
the vocabulary specificity collapse
concrete nouns carry voice. "the 11pm slack from a founder who'd just lost his cto" is specific. "a stressful business situation" is accurate but generic. the specific version sounds like a person. the generic version sounds like a model summary.
we measure this as a ratio: concrete nouns divided by total nouns. strong voice profiles in our database run above 0.65. drifted writing drops below 0.40. ai-generated first drafts land in the 0.30 to 0.40 band almost without exception.
the founder with the lost cto is why you had something to say. the generic stress situation is what remains when that specificity gets smoothed out.
what to check before you publish
the fix isn't to use less ai. the fix is to notice what changes and correct the specific drift.
- score your sentence-length variation. run any draft you didn't fully write yourself through a profile check. you want the standard deviation number, not a pass/fail verdict.
- scan for your missing transitions. ctrl-f for the three phrases you use most often. if you find them in fewer than half your recent posts, the drift is structural.
- check vocabulary specificity. look for five concrete nouns that survived the draft. if you can't find five, the piece has lost its particularity even if the facts stayed accurate.
- write your own opening. the first three sentences carry more voice signal than anything else in the piece. never let ai draft the opening. this single habit prevents most of the drift from accumulating.
ai detection flags what collapses first. hold your voice tracks your voice profile across every draft so you see the drift before it accumulates.
start free trial →
writes about brand voice, ai writing patterns, and the craft of sounding like yourself. built hold your voice after watching his own voice flatten across six months of heavy ai drafts.
co-written with ai as sidekick. shashank drafted the observations; the ai pressure-tested the structural claims. if something reads too smooth, that's the ai's fault.