I had to make this one a video: Substack’s detector repeatedly called my recursive gestalt writing—and even a transcript of my own speech—100% AI. The absurdity became the experiment, and the experiment became the story.

After completing the series, Why GLP Writers Write So Differently, I thought I had finished examining the recursive architecture of gestalt writing—the way meaning begins as a whole and language arrives later, gradually finding forms capable of carrying it. Then Substack introduced an AI detector, and the series unexpectedly became an experiment. When I submitted an entirely human-written autobiographical essay about my own experience as an autistic gestalt language processor, the detector classified it as one hundred percent AI.

Drawing on more than two decades in digital and multimedia forensics, I began testing the result rather than merely objecting to it. I changed punctuation, sentence structure, spelling, tone, formatting, subject matter, and method of entry. I compared the essay with older forensic articles of mine and with literary writing by Virginia Woolf, all of which the detector readily recognised as human. Yet every version retaining my contemporary recursive, autotheoretical register continued to be classified as entirely artificial.

The clearest contrast appeared when the detector accepted a contemporary statement in which I wrote directly to Notes that argued about disability, assistive technology, and authorship. It recognised me when I established a position, explained the stakes, anticipated objections, and defended a conclusion. What it repeatedly rejected was the register that returns rather than advances—the writing that develops through recursion, relation, reclassification, metaphor, and the quarter turn by which an earlier truth remains true but becomes insufficient.

The finding became more significant when I tested an unedited transcript of myself speaking. The transcript contained repetitions, false starts, transcription errors, self-corrections, and visible human disfluency, yet it too was classified as one hundred percent AI. The detector was therefore not simply responding to polished prose or literary smoothness. Across both written and spoken language, it appeared to be responding to a particular organisation of meaning—a reproducible, cross-modal false positive supported by a within-author contrast.

This does not prove that AI detectors are useless, that their published benchmarks are false, or that all autistic or gestalt writers will be misclassified. It does show that validation claims have boundaries, and that rare forms of human discourse may fall outside the populations and registers represented in benchmark data. When a statistical classifier is presented as an instrument of authentication, probability becomes an identity judgement: the machine produces a score, and the disabled writer must defend her tools, her authorship, and eventually the reality of her own mind. The central question is therefore not whether technology touched the writing process, but whether our systems recognise the full range of ways human beings organise meaning.

The AutSide is a reader-supported publication. To support my work, consider becoming a paid subscriber.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit autside.substack.com/subscribe