i keep seeing ‘perplexity’ and ‘burstiness’ mentioned when people explain how detectors work but nobody explains what they’re measuring precisely. what does each signal capture and why do detectors weight them?
perplexity measures how surprising the next token is given the previous ones. a language model assigns a probability to every possible next word. human writing chooses surprising words more often than AI, which tends toward high-probability continuations. low perplexity means predictable word choices.
burstiness measures sentence length variation. human writing naturally alternates between short and long sentences. AI produces more consistent sentence lengths within a passage. that consistency is measurable and distinct from human writing patterns.
the reason these two signals matter more than vocabulary: they’re harder to fake deliberately. you can swap words consciously. varying sentence rhythm while writing naturally is harder to maintain. which is why tools that target burstiness specifically tend to move detector scores more than paraphrasers.
token probability is the more granular version of perplexity. instead of measuring surprise across a passage, it can flag specific tokens where the model’s confidence that a human would choose that word is low. some tools return token-level heat maps showing exactly where the text diverges from expected human patterns.
synthesizing from the replies: the three signals work together.
perplexity: how predictable is each word choice across the passage? AI picks high-probability words. human writing is less predictable.
burstiness: how much does sentence length vary? human writing has high variance. AI writing has low variance.
token probability: the per-token version of perplexity, enabling flagging at the sentence or phrase level rather than just the document level.
tools that target these signals specifically (like Walter Writes, free lite without signup, 300-word trial on signup, no card) are doing something different from paraphrasers. paraphrasers change vocabulary. humanizers that target perplexity and burstiness change the underlying statistical profile of the text.