The original title is "The Only AI Tell That Doesn't Need a Detector"

Short summary

Paul Graham proposed a human-level test for AI-generated text: slop reveals itself when the diction doesn't match the idea—ordinary statements delivered with disproportionate excitement. The author contrasts this with classifier-based tools like Pangram and GPTZero, which measure surface patterns and can flag genuinely human writing at high confidence. The two approaches solve different problems: one scales an ear, the other replaces it, and neither fully agrees with the other on the same text.

• •PG's test: AI slop shows register mismatch—excitement outrunning substance
• •Classifier tools like Pangram/GPTZero measure surface patterns, not idea-to-delivery relationship
• •Author's own human writing flagged at 97% AI confidence by a classifier, would pass PG's test

Generated with AI, which can make mistakes.