Amplifying, Not Learning: Fine-Tuned AI Text Detectors Amplify a Pretrained Direction
A new research paper suggests that AI text detectors do not learn to distinguish between AI-generated and human-written text. Instead, these detectors amplify a pre-existing directional bias in their training data, effectively creating a 'typicality' axis rather than a true AI-vs-human boundary. The study found that raw, unfine-tuned encoders often perform as well as or better than fine-tuned detectors, and that the same axis can be inverted when applied to non-native English writing. AI
IMPACT Challenges the effectiveness of current AI text detection methods, suggesting a need for re-evaluation of their underlying mechanisms and potential biases.