DeepMind and UPenn make AI-text detectors auditable with sparse autoencoders
rohanpaul_ai · x · 2026-10-23
A Google DeepMind + UPenn paper (EMNLP 2026, arXiv:2610.30006) converts black-box AI-text detectors into interpretable classifiers:
- Train sparse autoencoders on the detector's residual stream across 4 layers, splitting activations into readable features;
- Auto-label each SAE feature and mask out those failing a global fidelity threshold;
- Train a lightweight logistic regression classifier on the remaining interpretable features.
The resulting classifier matches the black-box original and outperforms both neural and stylometric baselines out-of-distribution. Extracted features map to specific AI "tells" — em dashes and "delve into" push toward AI, "Firstly," pulls the other way. The authors argue detectors used in peer review, education, admissions, and moderation should expose and allow removal of the cues they rely on.
More from Safety
- Florida AG asks court to bar OpenAI from training new models without third-party approval — imjustnewatai · 2026-10-23
- OpenAI to join Prompt2PWN 2026 AI hacking conference for live vuln research — OpenAIDevs · 2026-10-23
- Science lifts expression of concern on 2024 AI debunking-conspiracies paper; conclusions unchanged — DG_Rand · 2026-10-23
- Chamath: Banks will block AI agents via terms of service before anything else — nicolechirps · 2026-10-23
- Reddit thread argues safety research is indistinguishable from capabilities research — ResultBackground2450 · 2026-10-23
- Beam vs GLM 5.3 Abliterated: 10-round attack/defense test shows 96% of vulns patched — bigblueboo · 2026-10-23