Paper summary
Paper summary: Feature-Level Insights into Artificial Text Detection with Sparse Autoencoders
Using sparse-autoencoder features from Gemma-2-2B's residual stream to detect and explain AI-generated text, revealing a distinct LLM writing style, especially in information-dense domains.
- Sparse autoencoder features turn "this text looks machine-written" into specific, inspectable features.
- No single detector works reliably across unseen text types and new LLMs, which is why interpretability matters here.