DeepMind open-sources DetectorScope, an AI text detector that explains its verdicts feature by feature
algo_diver · x · 2026-10-29
Google DeepMind and UPenn released DetectorScope (EMNLP 2026), open-sourcing code and models. It turns black-box AI text detection into auditable predictions that are sums of named, located features (e.g. preposition "about" +0.23). F1 beats Binoculars by 8-12 on unseen text with little accuracy loss from explaining itself. GPT-4o text showed telltale "::" after bold headers across 54k samples. Caveat: feeding the feature list to Gemini let it evade detection 60-100% of the time with under 40% rewriting. Not intended for high-stakes individual judgments.
More from Research
- Axiom's agents are formalizing CFSG in 43M lines of code, possibly the largest open-source project ever — SucceededMind · 2026-10-29
- MIT's 'AI telepathy' KV-cache transfer beats text by 40% accuracy at 4x speed, across model vendors — LucaAmb · 2026-10-29
- Coding agents fail to reimplement models from papers, sparking doubts about how papers are written — charlieharris01 · 2026-10-29
- Information-theoretic scaling laws for diffusion models predict when memorization begins — davidad · 2026-10-29
- Amandine Flachs launches side project orbench — IgorCarron · 2026-10-29
- Critic says closed weights hide personas from users, but that doesn't prove models aren't general-purpose actors — aiamblichus · 2026-10-29