Google DeepMind's IdeaLens reads outlines to detect AI-generated ideas, not AI-written words
rohanpaul_ai · x · 2026-10-07
A new paper from Google DeepMind and other top labs proposes IdeaLens, a detector for AI-generated ideas rather than text. Reading a document's outline instead of its words, IdeaLens flagged only 7% of AI writing from human outlines (Pangram 4: 92%), but caught 68% of human stories written from AI outlines (Pangram 4: 8%). The finding: wording-based detectors judge who wrote the words, so rules about original thinking can't rely on them — IdeaLens was trained on paraphrased point lists so wording couldn't help it.
More from Safety
- Ex-OpenAI/Anthropic researcher: automated AI researcher by 2027-28, humans far out of the loop — rohanpaul_ai · 2026-10-07
- AI safety researcher David Krueger calls for immediate international moratorium on frontier AI — DavidSKrueger · 2026-10-07
- Strict liability for AI outperforms slowdown or pre-release authorization, panel finds — Miles_Brundage · 2026-10-07
- Why a 'Butlerian Jihad' makes one AI watcher put a low P(doom) — johnohallman · 2026-10-07
- ChatGPT Flags Family Beach Video of Kids as NSFW Child Porn — FormalAd7367 · 2026-10-07
- AI is changing the security debt equation, researcher warns — WeldPond · 2026-10-07