Substack's new AI detector sparks a writer 'witch hunt'
Substack rolled out a Pangram-powered tool that scores how much of a post looks AI-written. The people it's meant to protect, warning that one false flag can, as one put it, 'irreversibly' and tarnish a reputation. Pangram claims a false-positive rate of about one in 10,000; spread across the millions of posts on Substack, that is still a lot of wrongly-accused people.
What it actually means
The lesson is about wiring a probabilistic AI classifier into a decision that lands on a real person. It could be in the form of a ban, a flag, a rejected application. A system that's 99.99% right sounds bulletproof until you run it at scale and count the wrong calls in absolute numbers. If you're building automated judgement into a product, the question isn't 'how often is it right' but it's 'what happens to the person when it's wrong, and who checks before it does.' Keep a human on the exceptions.
A 99.99%-accurate classifier still wrongs real people at scale.
Source: eWeek · Sunday 2 August 2026