Original post · Original
The article text is unavailable in this language; an existing version is shown.
Rohan Paul@rohanpaul_ai
A black-box AI-text detector can be rebuilt from readable parts without losing its strength.
Each verdict then comes with weighted writing cues a reviewer can check, like em dashes pushing toward AI and “Firstly,” pulling slightly the other way.
activations from 4 layers of a fine-tuned AI-text detector are split into features by sparse autoencoders, features without reliable labels are masked out, and a logistic regression classifier makes the call from what remains.
Title: "Building Auditable AI Text Detectors with Sparse Autoencoders"
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.
View this post on X ↗
· UTC+8