Google DeepMind and Penn NLP researchers have released DetectorScope: open code and models that take a strong but opaque AI-text detector and rebuild it so each prediction is a sum of human-readable features. The paper, *Building Auditable AI Text Detectors with Sparse Autoencoders* (arXiv:2610.30006), was submitted on October 21, 2026. On October 25, 2026, the project README News section said DetectorScope will appear at EMNLP 2026 in Budapest—that is a conference appearance note, not a claim that the work first launched that day.
What the pipeline does
The pipeline trains sparse autoencoders (SAEs) on a detector's residual stream, assigns each feature a short natural-language description, keeps only features whose descriptions pass a global validation score, and trains a lightweight classifier on those features. A call can return a label, a probability, and a ranked list of features with importance scores; optional highlights show where in the text each feature fired. Models cover open-domain COLING-style detection and closed-domain peer-review detection. Install with `pip install detectorscope`; the README links to a Hugging Face Spaces demo.
Why a readable score is still not proof of cheating
That explainability does not turn a score into proof of cheating. The repository Disclaimer states that the detectors are research artifacts; that they make mistakes—including on text written by people; that accuracy outside the domains and generators they were trained on is not established; that they must not be used as the basis for decisions about individuals, such as academic integrity, hiring, or publication decisions; and that this is not an official Google product. The authors and Google accept no liability for any use of them.
Generation-side marking and detection-side explanation are two sides of the same problem. For the watermark angle, see OpenAI textGrain: Invisible Watermarks in the EU. We have not run the DetectorScope demo or verified claimed language coverage; figures and install steps above are for reference only. Feedback can go through the site-owner contact link in the footer.