DetectorScope's paper was submitted October 21. The public repository documents code, models and a demo; its October 25 news entry announces a forthcoming EMNLP appearance without dating those resources' release.
The big change
- What changed: Named features explain the classifier's score.
- Why it matters: Researchers can inspect questionable formatting cues behind classifications.
- What to watch: The explanation cannot establish authorship. The repository rules out individual decisions based on these research detectors.
Method
Sparse autoencoders extract features from a language model's activity. Generated descriptions are checked against feature activations; poor matches are excluded. A sparse classifier reports how each remaining feature contributes to the score. For input text, the documented API returns a label, probability, feature contributions and optional token highlights.
Evidence
The study tested English benchmark texts and peer reviews, including held out generators. Targeted paraphrases generally changed scores as predicted; repeated guided edits flipped many labels.
Descriptions can miss activation conditions, and the study used one Gemma based model family. The repository cautions that its research detectors err on human text and must not decide individual cases.
BIG CHANGE's earlier watermark coverage examined a different kind of text-origin signal. The repository documents a Python package, downloadable models and a demo. BIG CHANGE has not run them or verified deployment cost. Individual authorship requires evidence beyond a detector score.
