Brilliant paper on interpretable AI-text detection from Google DeepMind and UPenn.

The strongest AI-text detectors are fine-tuned classifiers that give no reason for a flag, and the readable stylometric alternatives are much less accurate. That gap matters when a flag affects a peer review, an admission, or a student.

This work trains sparse autoencoders on a detector's residual stream, labels each feature automatically, and keeps only features whose descriptions pass a fidelity check. A small classifier is then trained on those features alone.

The kept features correspond to specific linguistic markers of AI-generated text. The interpretable classifier also outperforms both the original neural detector and the stylometric baselines on out-of-distribution text.

Accepted at EMNLP 2026, with code and models released.

https://arxiv.org/abs/2610.30006

Chat with Paper: https://academy.dair.ai/papers/building-auditable-ai-text-detectors-with-sparse-autoencoders-2610.30006