| title | Building Auditable AI Text Detectors with Sparse Autoencoders |
|---|---|
| authors | Liam Dugan, Callum McDougall, Matthew Denton, Philip Pham, Christine Kaeser-Chen, Chris Callison-Burch |
| date | 2026-10-21 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2610.30006 |
| affiliation | GoogleDeepMind |
| upvotes | 14 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf code |
| abstract | AI-generated text detectors are increasingly deployed in consequential settings---peer review, education, admissions, journalism, and content moderation---where their outputs directly affect individuals. Yet the strongest detectors are fine-tuned neural classifiers whose decisions are opaque, while transparent alternatives (stylometric and linguistic-feature classifiers) underperform them by wide margins. We present a pipeline that converts a fine-tuned neural detector into a classifier that operates on human-readable concepts. We accomplish this by training sparse autoencoders (SAEs) directly on a detector's residual stream, assigning auto-interpretation labels to each SAE feature, and retaining only features whose generated descriptions pass a global fidelity threshold. We then train a lightweight classifier on just the interpretable features. We find that our SAEs recover latent concepts that correspond to highly specific and predictive linguistic "tells" of AI-generated text. Classifiers built on these extracted features are not only highly interpretable but exhibit stronger performance on out-of-distribution texts than either neural or stylometric alternatives. |
cite
@misc{dugan2026building,
title={Building Auditable AI Text Detectors with Sparse Autoencoders},
author={Liam Dugan and Callum McDougall and Matthew Denton and Philip Pham and Christine Kaeser-Chen and Chris Callison-Burch},
year={2026},
eprint={2610.30006},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2610.30006}
}