Skip to content
Trending storyDeveloping

Google DeepMind and UPenn build an interpretable AI-text detector matching a black-box one

2 articles2 sourcessince Oct 22Last article 7h ago ·

Overview

AISummary of 2 articles

Researchers from Google DeepMind and the University of Pennsylvania trained sparse autoencoders on a Gemma-2-2B AI-text detector, labeled its features, and kept only those whose labels predicted behavior well.

The resulting detector explains its verdicts with readable writing features. In-domain, it scored 96.8 F1 against 98.0 for the black-box detector. On unseen test data, a version built on base-model features scored 92.2 F1 against 75.2 for the black box, according to the paper, which Rohan Paul shared on X.

A later X post names the method Detector Scope and says the paper is accepted to EMNLP 2026. That post describes the method as converting black-box AI text detectors into auditable models, and it examines which features detectors rely on when identifying generated text.

Written by AI from the articles below · updated Oct 29, 10:07 PM ET

Check the sources:

Latest developmentOct 29, 4:01 PM ET
Detector Scope uses sparse autoencoders to audit AI text detectors

Article timeline

The articles in this story. Times are ET.

Oct 29
  1. Liam DuganX
    Detector Scope uses sparse autoencoders to audit AI text detectors

    AIResearchers with Google DeepMind introduce Detector Scope, a method that uses sparse autoencoders to turn black-box AI text detectors into auditable models. The paper, accepted to EMNLP 2026, examines what features AI detectors rely on when identifying generated text.

    GIF from @LiamDugan_'s post
Oct 22
  1. Rohan PaulX
    Google DeepMind and UPenn build AI-text detector with readable, auditable features

    AIResearchers from Google DeepMind and the University of Pennsylvania trained sparse autoencoders on a Gemma-2-2B detector, labeled its features, and kept only those whose labels predicted behavior well. The resulting detector scored 96.8 F1 in-domain against 98.0 for the black box, and 92.2 F1 on unseen data against 75.2 for the black box.

    Image from @rohanpaul_ai's post

Heat Trend

Current comparable heat
8
Peak comparable heat
8Oct 29, 11:00 PM ET
Change in past 24 hours
–

The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.