Substack's New AI Detector Has the Same Blind Spot DEV.to's Did

Short summary

Substack launched an AI detector powered by Pangram, but it inherits the same false-positive problem that plagued DEV.to's manual moderation: well-written, formal, or non-native English text gets flagged as AI-shaped. The author argues detectors catch writing patterns, not whether human thinking went into the work, creating a dishonesty incentive where undisclosed AI use goes undetected while transparent use gets penalized.

• •Substack's Pangram-based AI detector shares the same false-positive blind spot as DEV.to's manual GPTZero moderation
• •Formal, well-structured, and non-native English writing is disproportionately flagged as AI-generated
• •A mostly-reliable detector can be more dangerous than an obviously unreliable one because people stop checking

Generated with AI, which can make mistakes.