The authors first fine-tuned a neural model to distinguish AI-generated text from human writing.
AI text detectors can expose readable features without sacrificing performance
Dugan and colleagues report that classifiers built from interpretable internal features preserve detection performance and outperform neural and stylometric alternatives on out-of-distribution text.
Big Tech
Liam Dugan · Callum McDougall · Matthew Denton · Philip Pham · Christine Kaeser-Chen · Chris Callison-Burch
University of Pennsylvania · Google DeepMind
Research Digest··2 min read
Dugan and colleagues present a pipeline that turns a neural AI-text detector into a classifier using features with natural-language descriptions.
§