everyone's been flagged at some point. you write something yourself, run it through an ai detector, and it comes back 60% ai-generated. feels great.

so what's actually going on under the hood?

the two things detectors measure

it mostly comes down to perplexity and burstiness. fancy names for simple ideas.

perplexity is about word predictability. if you write "the cat sat on the ___," most people would say "mat." ai would too. low perplexity means the text is predictable. ai text tends to be very predictable because it literally picks the most likely next word.

human writing is messier. you might write "the cat sat on the radiator" or "the cat sat on my laptop again." those are less predictable. higher perplexity. detectors notice.

burstiness is about rhythm. read any essay you've actually written yourself. some sentences are long and winding. some are two words. maybe you threw in a parenthetical that didn't really need to be there. ai doesn't do that. it writes in this steady, even cadence — sentence after sentence of roughly the same length and complexity.

where they break down

here's the thing detectors don't advertise: they're probabilistic. they're making a guess.

non-native english speakers get flagged constantly because they tend to use simpler, more common vocabulary. that looks like low perplexity to a detector. technical writing gets flagged because it's formulaic by nature. if you're writing a methods section for a research paper, yeah, it's going to sound uniform.

gptzero's own studies show false positive rates around 2-3%. sounds low until you realize that's thousands of students getting wrongly accused every semester.

what this means for you

if you're using ai as a writing tool — brainstorming, drafting, getting unstuck — the output will absolutely get flagged if you submit it raw. the patterns are too clean.

that's where untraceable ai tools come in. not synonym swappers. actual rewriters that restructure sentences and add natural variation. paraai's paraphrase tool was built specifically for this — it breaks up the patterns that detectors rely on.

the detector arms race isn't slowing down. but understanding how they work is the first step to untraceable ai writing that actually holds up.