Building Auditable AI Text Detectors with Sparse Autoencoders
Date:
Abstract
In this talk I discuss work from my internship at Google DeepMind on building auditable versions of existing black box AI text detectors. Our method is fairly simple. We train sparse autoencoders on the residual stream of the fine-tuned models, filter for only the most interpretable features, then learn a logistic regression head on top of the new feature set. We show that these auditable classifiers retain the generalization performance of the underlying black box model while simultaneously being far more interpretable.
Location
This talk was recorded on October 26th, 2026 as part of the EMNLP 2026 conference in Budapest, Hungary.