Breaking01 / AI
Auditable AI text detector reports small drop in performance
The authors say the detector must not be used for decisions about individuals and that its accuracy beyond its training scope is not established.

Google DeepMind and University of Pennsylvania researchers reported in Building Auditable AI Text Detectors with Sparse Autoencoders that DetectorScope loses little performance on the RAID benchmark (5% false-positive rate); no independent replication is confirmed.
The repository claims that, unlike black-box detectors, every prediction is a sum of validated features. A sparse autoencoder splits a language model's representations into features, each with a description and validation score. Only those at 0.85 or above (default, logistic-regression classifier) feed the classifier.
The open-domain model uses COLING 2025 GenAI Detection shared-task data (40 generators). The authors reported scores near the base black-box classifier on COLING F1 and RAID detection rate.
A repository notice says the detector also errs on human-written text and must not underpin decisions about individuals, such as academic integrity or hiring. Reproduction code is public; no outside results are listed.
Accuracy beyond trained domains and generators is not established, the authors wrote. The paper is due at EMNLP 2026 in Budapest.