AI detectors flag pre-ChatGPT writing as up to 95% AI-generated
A writer ran their own writing from 2019 onward — all predating ChatGPT — through 11 different AI-detection tools. The results were bad: some tools rated as little as 5% of the text as human-written, despite it being entirely human-authored with no AI involvement whatsoever. This happens because these detectors judge text based on stylistic and structural patterns, such as unusually smooth or consistent phrasing, meaning that anyone who naturally writes in a polished or regular style risks being wrongly flagged as an AI author.
Key points
- The writing tested spans 2019 onward, entirely before ChatGPT existed
- 11 separate AI-detection tools were used in the test
- Some tools scored as little as 5% of the text as human-written
- The results highlight how prone AI detectors are to