The ship of Theseus paradox in AI-assisted writing

"They wrote something genuine, then passed it through an AI writer of their choice ten times in a row. After each iteration, they rated it on a 1-7 scale: "How much is this still mine?" The survey itself ran on Qualtrics; I cleaned and visualized the data in R, using a simple mixed-effects model to trace how that sense of ownership decayed over time. Everyone consented, all responses were anonymized, and what's shown here reflects only the aggregate picture."

"Perceived ownership steadily declines with each AI rewrite, dropping fastest for texting and slowest for code. Turns out, the ownership ratings fell off a cliff around the third or fourth rewrite for texting and essay writing, but held surprisingly steady in code. A spline model (R, lme4, three-knot natural spline) showed the steepest drop early on: texting collapsed fastest, essays followed, code barely flinched."

"This aligns with a broader cultural confusion around what "human-written" even looks like. As Aaron Pace recently wrote, the more polished a piece of writing is, the more likely AI detectors are to label it "machine-made." Today we see a strange modern reversal where professionalism now reads as artificial. My data shows the inverse feeling: the messier and more intimate the writing, the more its owner wants to keep machines out."

Participants created original content in three modes—casual text, short essay paragraph, and code snippet—and applied ten sequential AI rewrites. After each iteration, participants rated perceived ownership on a 1–7 scale. Responses were collected and analyzed with mixed-effects models and a three-knot natural spline to trace ownership decay across iterations. Ownership declined steadily, most sharply for texting, moderately for essays, and minimally for code. Texting ownership fell below midpoint around the third rewrite, essays around the seventh, while code rarely crossed that threshold. Intimate, messy writing resisted AI adoption more than polished professional code.

Read at Medium