Paul Graham's AI Detection Test Targets Tone-Idea Mismatch, Not Just Text Patterns

Paul Graham argued this week that AI-generated 'slop' reveals itself when the excitement of the writing outpaces the ordinariness of the idea being expressed — a mismatch in register rather than a statistical surface pattern. This contrasts with tools like Pangram, integrated into Substack by CEO Chris Best, which detect AI writing by scoring token-level and sentence-level patterns against known machine output. A writer on DEV.to tested both approaches on the same pieces and found they disagreed: classifier tools flagged substantive, plainly written human work at 97% confidence, while Graham's intuition-based test would have cleared it. The author notes that a non-native English speaker who spent up to 20 hours crafting an argument and then polished phrasing for flow does not produce the register mismatch Graham describes, yet automated tools flagged the post regardless. The piece concludes that the two detection methods solve different problems — one scales human intuition across a fast-moving feed, while the other requires a careful reader who knows what to listen for.