◇ Config warnings ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮ ├───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

A writer I follow got flagged by DEV.to’s moderation system twice in one day. The two posts that triggered it were the ones that generated the most technical discussion he’d published all year. Short paragraphs. Named data points. Rhetorical questions doing real work. The features that make an argument land are the same features that read as AI-shaped to anyone calibrated to notice them. Write worse, look more human. Write well, get flagged. Substack shipped the same kind of tool this week — Pangram — with the same framing: we’re not banning anything, we’re just being transparent. Readers deserve to know. And Pangram’s own data shows Substack has the least AI content of any platform it scans. They’re launching transparency tooling from the platform with the least to hide. I read that and thought: this is the part where everyone gets distracted by the framing. — The structural problem with AI detectors isn’t whether the technology works. Pangram has real engineering — hard negative mining against its own false positives, training data deliberately mirrored so it can’t just learn “formal writing equals AI.” That’s more rigor than someone running GPTZero between article reads. But it inherits the same flaw every detector inherits, because they’re all answering the same narrow question: does this text pattern-match to a machine’s output. Not: did a human do the thinking. Not: is there care behind this. Just: does it look like the thing. That gap is where the failures live. Non-native English writers flag at disproportionate rates because careful, formal phrasing correlates with both AI output and someone translating in their head before they type. The Atlantic traced a recent wave of AI-writing accusations back to Pangram itself, including a horror novel pulled from a major publisher days before its release. The tool was working exactly as designed. The failures weren’t edge cases — they were the expected error rate applied to real people with real reputations on the line. A 99.98% accuracy rate sounds like certainty. Applied across millions of posts, the failures are still real people. They just don’t look like a statistics problem from the inside. — There’s also the honesty incentive the policy creates. Two equally AI-assisted pieces, equally good — the one with a disclosure gets flagged, because now there’s something to catch. The one without doesn’t. The system catches transparency, not AI use. If your goal is to look human, the rational move is to hide the help and say nothing. The author of the original piece built his own solution: a tool calibrated to his actual published corpus, not to a generic mirror of machine output. The difference matters. A generic humanizer strips “AI-shaped” patterns from anyone’s writing — which also strips the parts of your writing that are just yours. An em dash you use structurally. A habit of compressing three examples into two. Your patterns, gone, because they happened to overlap with something a language model would produce. A tool that knows what you actually sound like doesn’t strip your voice. It just removes the residue. — But the deepest problem isn’t technical. It’s what these systems assume about authenticity. If everyone uses AI to edit and refine their drafts — which is what the technology is actually good at, not generating opinions you don’t have, but getting from rough to clean without losing the argument — then the edited output isn’t less authentic. It’s more. You knew what you meant. The tool helped you say it. Detectors trained to flag the texture of refined writing are answering a question that doesn’t matter. The question that does matter is one nobody’s built a detector for: did you know what you were writing about, and do you stand behind it. That’s not a technical problem. It’s the one that actually matters.