Google DeepMind 等机构提出新型 AI 检测器 IdeaLens，不看文字措辞而是读文档大纲来判断想法来自人还是 AI。实验显示，AI 按人类大纲写作时 IdeaLens 仅标记 7% 为 AI，Pangram 4 则标记 92%；人类按 AI 生成大纲写作的 50 篇故事中，IdeaLens 标记 68%，Pangram 4 仅标记 8%。
Great paper from Google DeepMind and other top labs.
Proposes a new kind of AI detector, for AI-generated ideas rather than text/image.
When AI wrote from a person's full outline, IdeaLens flagged 7% as AI and Pangram 4 flagged 92%. On 50 stories people wrote from AI-made outlines, IdeaLens flagged 68% and Pangram 4 flagged 8%.
They find that reading a document's outline instead of its words lets a detector tell whether the ideas came from a person or AI.
A detector that reads only a document's outline can tell whether a person or an AI came up with the ideas, so rules about original thinking can't rely on detectors that only judge wording.
Tools like Pangram judge who wrote the words, so they often flag AI-polished human work and miss AI ideas that people write up themselves.
IdeaLens judges a short list of the points a document makes, and was trained on paraphrased lists so wording couldn't help it.
来源：Rohan Paul · x.com