<!-- bryant-content-generator:start --> # This document is also an excellent demonstration... > This document is also an excellent demonstration of how completely AI-writing detectors can misunderstand authorship. Published evaluations have found that detection rates can fall sharply under paraphrasing, that human prose can be falsely classified, and that some detector methods systematically penalize non-native English writing. This document may nevertheless be labeled “AI generated” because a classifier recognizes statistical features of its present surface. As an account of where the work came from, that conclusion would be preposterous. The present prose may be AI-assisted, but AI did not go back to 1994 and build the symbolic language engine, conduct the following decade of natural-language and informatics research, begin the Atomic Markdown system around 2000, deploy its publishing and syndication machinery, or spend twenty-six years refining the authorial practice this project explains. An AI detector has no time machine: it cannot go back and perform the twenty-six years of original engineering, deployment, and revision whose history the document contains. {{Detector classification is not authorship analysis; it is pattern recognition performed after the causal history of the work has been discarded.}} [[reminders/Information/Detector Classification Is Not Authorship Analysis by Bryant McGill|∴]] **Source:** [[projects/The Argument in Miniature - How I Write Quotations, Paragraphs, and Reusable Prose|The Argument in Miniature: How I Write Quotations, Paragraphs, and Reusable Prose]] ## Share on Social Media ``` “This document is also an excellent demonstration of how completely AI-writing detectors can misunderstand authorship. Published evaluations have found that detection rates can fall sharply under paraphrasing, that human prose can be falsely classified, and that some detector methods systematically penalize non-native English writing. This document may nevertheless be labeled “AI generated” because a classifier recognizes statistical features of its present surface. As an account of where the work came from, that conclusion would be preposterous. The present prose may be AI-assisted, but AI did not go back to 1994 and build the symbolic language engine, conduct the following decade of natural-language and informatics research, begin the Atomic Markdown system around 2000, deploy its publishing and syndication machinery, or spend twenty-six years refining the authorial practice this project explains. An AI detector has no time machine: it cannot go back and perform the twenty-six years of original engineering, deployment, and revision whose history the document contains. Detector classification is not authorship analysis; it is pattern recognition performed after the causal history of the work has been discarded. ∴” — Bryant McGill https://bryantmcgill.com/passages/passages-document-excellent-demonstration-history-work-discarded ``` <!-- bryant-content-generator:end -->