Building Auditable AI Text Detectors with Sparse Autoencoders
The authors report that detectors using human-readable features outperform neural and stylometric alternatives on out-of-distribution text.
The pipeline trains sparse autoencoders on a neural detector’s internal representations to extract linguistic concepts. It automatically labels those features and retains only those whose descriptions pass a fidelity threshold. A lightweight classifier then uses the retained features to detect AI-generated text. The authors say these features reveal specific, predictive linguistic “tells,” making the classifier’s decisions easier to audit.
ArXiv · AI/CL/LG's note
The pipeline trains sparse autoencoders on a neural detector’s internal representations to extract linguistic concepts. It automatically labels those features and retains only those whose descriptions pass a fidelity threshold. A lightweight classifier then uses the retained features to detect AI-generated text. The authors say these features reveal specific, predictive linguistic “tells,” making the classifier’s decisions easier to audit.
ArXiv · AI/CL/LG's note
score 4