Researchers from Google DeepMind and the University of Pennsylvania developed a method to make AI text detectors interpretable. The approach uses sparse autoencoders to translate dense neural network signals into human-readable concepts, allowing users to see the logic behind classification decisions.
The pipeline utilizes a fine-tuned Gemma-2-2B model. Sparse autoencoders decompose internal residual stream activations into individual features, which are then automatically labeled. A logistic regression classifier uses these features to determine if text is AI-generated, providing weightings for each factor.
On in-domain data, the interpretable model achieved an F1 score of 96.8, compared to 98.0 for the original black-box detector. On out-of-distribution data, the new model reached an F1 score of 92.2, significantly outperforming the baseline score of 75.2.
The research, titled Building Auditable AI Text Detectors with Sparse Autoencoders, was submitted on October 21, 2026, and accepted at EMNLP 2026. The source code and trained models are available in the google-deepmind/detectorscope repository on GitHub.
