# AI Editing Toward the Probable Is What Detection Detects ## Claim AI-based writing detectors trigger primarily on AI *editing* — the rearrangement of a writer's own words into a statistically more probable order — rather than on AI authorship, which means a person who writes their own content and has it AI-edited can be flagged while a person who generates content with AI and then roughens it is not. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for detectors built on next-word-probability (surprisal) scoring, which is how nearly all current commercial detectors work per the source. - Requires the editing pass to meaningfully reorder or restructure sentences, not merely correct spelling or grammar — a human editor who fixes coherence without rearranging sentences does not trigger the same effect. - Does not claim detectors never catch genuine AI authorship — only that editing, independent of who originated the content, is sufficient on its own to trigger detection, which breaks the assumption that a positive score indicates non-human origin. ## Evidence A colleague of Katie Robbert's ("Becca") had her own dictated, self-authored posts repeatedly flagged as AI, despite writing every word herself. The cause: she dictates and then has Claude tidy the transcript. The mechanism, demonstrated with a Yoda minimal pair — "For 800 years have I trained Jedi" rewritten by AI as "I have trained Jedi for 800 years" — is identical vocabulary in a more statistically ordinary order. Only the reordered version flags. > [!quote] > when an AI tool edits, it rearranges the words back into a statistically probable pattern The demonstration produced a genuinely perverse corollary: Robbert's own unedited rough draft scored as more clearly human than her polished, AI-assisted final version — meaning proving authorship under these tools can require publishing worse writing. ## Links - [[AI & the Future of Knowledge Work MOC]] — what detectors actually measure, and why authorship inference misreads it - [[3. Resources/Permanent Notes/Shared AI Tools Homogenize Output While Feeling Original to Each User]] — because contrasts; that note concerns general output convergence across users of the same model, while this note concerns a specific, measurable statistical mechanism (word-order probability) rather than felt originality. - [[3. Resources/Permanent Notes/A Defense Experts Fail Is Not a Defense for Users]] — because applies-in; if the underlying detection mechanism targets editing rather than authorship, a detector is measuring the wrong variable regardless of who is being tested with it, which strengthens the case against punitive use categorically. - [[3. Resources/MOCs/AI Agent Safety MOC]] — because applies-in; a concrete, testable failure mode for any system built on statistical detection rather than provenance. ## Tests / falsifiers - Run matched pairs of human-authored text (unedited vs. AI-edited-for-coherence-only) through a commercial detector and compare scores; the claim predicts the AI-edited version scores measurably higher regardless of who wrote the original content. - Test whether constraining AI editing to sentence-level reordering only (not sentence reconstruction) measurably preserves a human-like detector score relative to full AI rewriting — this is the source's proposed mitigation and is directly testable. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/So What How AI Detectors Work]] — reports the Becca case and the Yoda minimal-pair demonstration showing detection triggers on AI editing rather than authorship.