A simple test for AI detectors: separate pre-2020 from post-2020 human writing

RexDouglass · x · 2026-11-17

RexDouglass proposes an easily checkable experiment: add a head to an AI-text detector and have it distinguish pre-2020 from post-2020 human writing. How well it performs measures how much the detector is learning something beyond just "AI-ness"—i.e., general corpus-level vibes. This directly extends the thread's debate over whether older human text should be detectable.

Original post →

More from Models

Models channel →