Let's talk about Pangram, the tool Substack just partnered with for AI detection.

Every AI detector posts a false-positive rate from lab conditions. Then it meets real writing.

Turnitin went from a claimed 1% to 4% in the field. The messier the input - ESL, neurodivergent, heavily edited - the wider the gap between the brochure number and the person flagged.

Real world: these tools flag second-language writers at rates that hit 60%+ in independent testing. Black students get accused more often, controlling for everything else. The bias is structural, not a bug, a mathematical consequence of diverse populations that better engineering can't fix.

The instruments are racist. Racial bias, as outcome, lands as racism.

Substack is deploying this on every post published from today. The CEO of Pangram has been posting analyses of named journalists on Twitter/X calling it "receipts" - after previously publishing a disclaimer that it should never be used as a "sole arbiter" of anything. Disclaimer → exception

I said last week that AI detectors are notoriously unreliable (in yesterday’s podcast with Dan). I said just now this is a press gin-up.

Now I'm saying: this is not a neutral tool. The bias is documented. Deploy with extreme caution.