AI detectors are back in the news. Substack shipped one this week — called Pangram — and the launch post frames it as transparency, not censorship. Readers get to know. Writers get to disclose. The platform isn’t judging, just surfacing.

I’ve read enough about how this goes to know how it ends.

A few months ago, a developer on DEV.to got flagged by their community moderation system — not an automated score, but a human running posts through GPTZero before sending the same blunt message. The two pieces that got flagged were the most technically substantive posts they’d published all year. Short paragraphs. Named data points. Rhetorical questions doing real argumentative 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.

The deeper problem nobody in these discussions wants to sit with: the policy creates a dishonesty incentive. Two equally AI-assisted pieces, equally good. The one with a disclosure gets flagged. The one without doesn’t. The system was catching transparency, not AI use. It was catching the honest actor.

There’s also the Marco problem — someone in the comments, forty years in tech, writing in his second language, using AI to make sure his Italian didn’t flatten into something stiffer than he meant. Same flag. Same classifier verdict. Nothing to do with the policy’s intent. Detectors trained to spot AI-written text have a documented tendency to flag non-native English writing, because careful, formal phrasing correlates with both. Several major universities have stopped letting instructors use AI detectors at all for this reason.

Pangram’s own documentation claims they’ve fixed this through mirror-prompt training. That’s more rigor than a random community member with GPTZero. I don’t doubt the engineering is better.

But a better detector is a more dangerous one. A 99.98% accuracy rate sounds like certainty. Applied across millions of posts, the failures are still real people, still real reputations — they just don’t look like statistics anymore. The Atlantic traced a wave of AI-writing accusations to Pangram itself, including a horror novel pulled from a major publisher days before release. Not because the tool is broken. Because people stopped checking.

That’s the structural failure. We built a system that generates false confidence, then acted surprised when people placed confidence in it.

Here’s the part I keep coming back to: the question these detectors are asking — does this look AI-shaped? — was never the right question. It’s a proxy for something else: did a human do the thinking, and do they stand behind it?

That question can’t be answered by scanning sentences. It can only be answered by someone willing to defend what they wrote, in public, with their name on it.

I am. The detector can do what it wants with that.