DeepMind and UPenn build auditable AI-text detectors with sparse autoencoders, matching black-box F1
rohanpaul_ai · x · 2026-10-23
A Google DeepMind + UPenn paper shows an AI-text detector can explain its verdicts via readable features while matching a black-box detector.
How it works
- Train sparse autoencoders on a fine-tuned Gemma-2-2B detector's internals, use an LLM to label each feature, keep only well-predicting labeled features, and feed them into a simple classifier.
Results
- 96.8 F1 in-domain vs 98.0 for the black box
- On unseen test data, a base-model-feature version scored 92.2 F1 vs 75.2 for the black box — better generalization
Implication: detectors affecting people (student essays, peer reviews) should be built on labeled features so reviewers can inspect and remove unfair cues.
More from Research
- Reversible computing veteran poses open problem: universal BARCS primitives from unbiased Josephson junctions — MikePFrank · 2026-10-23
- Arc Institute uses genomic language model to map 376M prokaryotic noncoding sequences into 5.14M clusters — tedgreenwald · 2026-10-23
- The rise of AI-assisted peer review: reviewers are making it into the appendix — abursuc · 2026-10-23
- Anthropic uses Claude to formally verify security properties in 100+ open-source libraries at RISC-V assembly level — AnthropicAI · 2026-10-23
- Agentic interviews filter misaligned models, keeping LLM teams 81-83% clean — shangbinfeng · 2026-10-23
- Researchers test 9 strategies to pick LLM teams, beating random selection by up to 36.1% — shangbinfeng · 2026-10-23