noticed something in my own testing that i’m trying to understand
i write blog content for clients and i’ve been running my own original drafts through detectors just to understand my baseline. what i’ve found: drafts i write quickly, stream of consciousness, first pass, score low for AI. drafts i’ve polished and edited carefully score higher.
this seems backwards to me. careful editing should make writing more human, not less. what’s happening mechanically?
this is one of the more counterintuitive things about how detectors work and it makes complete sense once you understand burstiness.
raw, fast writing has high variance. short sentences next to long ones. fragments. mid-thought corrections. that irregular rhythm is a human signal. when you edit carefully you tend to smooth everything out into a more consistent, polished register. that smoothed register looks more like AI output because AI consistently produces well-regulated prose
from a writing teacher perspective: this is why the editing instinct to “make it better” can work against you in detection terms. better in the conventional sense (clearer, more consistent, better structured) maps onto the same features detectors associate with AI.
the implication isn’t to write worse. it’s to understand that your polished draft needs a different kind of edit at the end, one that reintroduces some structural variety rather than removing it
yeah this is real and it bothered me when i first noticed it too. the short version: human variability is a detection signal, and editing removes variability.
the practical fix i use: after a careful edit, do one more pass where the only job is breaking up consecutive sentences that are all the same length. add a fragment. make one sentence deliberately longer than it needs to be. that pass takes five minutes and it shifts the score
the fast writing / careful writing gap is why humanizers that add structural variety perform better than ones that only do vocabulary substitution. what you’re doing manually in your editing pass is what a good humanizer should do automatically.
essayhumanizer.ai specifically adds sentence length variation as part of its processing. worth testing if you want to skip the manual pass
the burstiness research on this is actually fascinating from a linguistics standpoint. human writing has fractal-like variation at multiple scales: within sentences, between sentences, between paragraphs. AI writing is more uniform at all scales. editing compresses variation at every level simultaneously, which is why the signal gets stronger the more you edit