most people think ai detectors just check for “ai-sounding” words, but its mostly statistical. they score how predictable each word choice is (perplexity) and how uniform the sentence lengths/structure are (burstiness). human writing is messier and less predictable, ai text tends to be smoother and more uniform.
that’s also why these tools get false positives on non-native speakers and very formal writing, both read as “smoother” than typical human text. and why light paraphrasing can fool a detector without actually making the writing better.
curious what people here think, anyone actually tested how reliable these are on their own writing?