Google DeepMind and UPenn paper proposes an interpretable AI text detector
A new paper from Google DeepMind and UPenn proposes an interpretable AI text detector that explains its decisions using readable writing features while achieving accuracy close to black-box detectors. The research trains sparse autoencoders on the internal activations of a Gemma-2-2B detector, decomposes its activations into features, and has an LLM label each feature, keeping only features whose labels can reasonably predict their behavior as input to a simple classifier. Using only well-labeled features, the detector achieves an in-domain F1 of 96.8, close to the black box's 98.0; on unseen test data, the version based on base-model features achieves an F1 of 92.2 versus 75.2 for the black box.
Why read this
AI score 56
The paper uses sparse autoencoders to decompose the black-box detector's internal activations into readable features, keeping accuracy close while making the basis for decisions auditable.
Topics
Body
New Google DeepMind + UPenn paper shows that an AI-text detector can explain its verdicts with readable writing features and still match a black-box detector.
suggests that anyone deploying AI-text detectors should be able to read and remove the specific cues a detector relies on.
The best AI-text detectors are black-box classifiers, yet they flag student essays and peer reviews. Readable stylometric detectors exist but trail them by wide margins.
They trained sparse autoencoders on a Gemma-2-2B detector's internals to split its activity into features, then had an LLM label each one. They kept only features whose labels predicted their behavior well and fed those into a simple classifier.
With only well-labeled features, the detector kept 96.8 F1 in-domain, close to the black box's 98.0. On unseen test data, a version built on base-model features scored 92.2 F1 against 75.2 for the black box.
If your detector's decisions affect people, build it on labeled features so reviewers can see and drop cues they consider unfair.