2026 November 1
I stumbled on the discrepancy by accident, though the test that produced it was deliberate. After uploading a photograph of myself to several of the leading artificial-intelligence image detectors, the systems returned a near-unanimous verdict: ninety-nine percent certainty that the image had been generated. It wasn’t. A professional photographer had taken it.
The result was interesting enough that I continued. I submitted six photographs from photoshoots on which I was the Creative Director and photography years earlier. Five were classified, again at ninety-nine percent confidence, as synthetic. All six predated the public arrival of the tools that now claim to recognize their own kind.
I turned next to the text classifiers used by publishers, teachers, and universities. Passages of my own writing from 2020 and earlier, years before the current generation of large language models were readily available or available at all. One detector, Pangram, scored three separate pieces at one hundred percent, ninety-five percent, and ninety-two percent AI. The consistency was absolute. When I asked a language model to examine the flagged work and isolate the features that had triggered the judgments, the answer was surprising. Formal diction, careful revision, and conventional grammar and punctuation had produced the effect. In the model’s view, these habits of composition resembled its own.
AI score: 100% AI written.
Researchers have reached similar conclusions from another direction. A 2023 study found that seven detectors wrongly flagged sixty-one percent of TOEFL essays written by non-native English speakers as AI-generated, against roughly five percent for native writers. Mike Perkins, an academic-integrity researcher, has put the practical implication plainly: the tools are “not fit for purpose.” OpenAI itself discontinued its own detector after concluding that the accuracy was too low to justify continued use.
My writing habits that the system penalizes aren’t mysterious. They were formed in public-school classrooms where my teachers required repeated writing and offered little patience for uncorrected errors. The requirement continued through college, where my professors, several of them published authors, returned my early papers marked in a density of red that made clear I had a lot of work to do. The instruction had been practical rather than theoretical: take the time, correct the sentence, prefer the clearer word. None of it had been designed to mimic a machine that did not exist.
When I asked an AI what changes would reduce the likelihood of future flags, the suggestions converged on a single recommendation: greater authenticity requires poorer writing.
I can imagine the commentary that Shakespeare, Dostoevsky, and Hemingway might have offered on the exchange. The more useful book would be a revised edition of Orwell’s Coming Up for Air, retitled Coming Up for Slop, in which George Bowling returns not to the Lower Binfield of his childhood but to the language of 2026, only to discover that the ordinary markers of care have themselves become grounds for suspicion.
Human’s use of AI detection tools doesn’t only fail to distinguish human writing from machine writing; they quietly redefine which forms of human writing still count as human.
I call that use of AI: Human Slop.
READY TO EXPLORE THE FULL PROCESS?
Join our community on Patreon for complete articles, creative process breakdowns, minimalist frameworks, and downloadable tools — including reusable templates and behind-the-scenes access.