A team from Google DeepMind and the University of Pennsylvania has developed a method that turns an opaque AI text detector into an interpretable classifier without losing much accuracy. The system uses sparse autoencoders to extract concepts from the residual stream of the Gemma-2-2B model, label them automatically and feed them into a lightweight logistic regression classifier. On data similar to the training set, the interpretable classifier achieves an F1 score of 96.8, compared with 98.0 for the original detector. On out-of-distribution data, it reaches 92.2, while the black-box baseline drops to 75.2. The code and trained models are publicly available on GitHub in the google-deepmind/detectorscope repository.
Source: Read the original article
