原帖引用 · AI 翻译
– https://arxiv.org/abs/2611.26216
标题:“前沿模型只需一篇文本,就能揭示其作者是人类还是 AI”
马里兰大学的一篇新论文发现,如今前沿模型检测 AI 撰写文字的能力已与 Pangram 等经过训练的检测器相当,而且还能解释其推理过程。
仅凭一篇文本,就能说出其作者、撰写它的 AI 实验室或具体的人。
匿名同行评审和化名账号都基于一个假设:仅凭文字无法揭示你的身份。人们还曾认为,识别 AI 文本需要经过训练的检测器,不能只靠普通提示词。
作者给模型提供单篇文本和一个零样本提示词,不允许网络搜索、使用工具,也不提供候选名单。他们测试了模型能否判断文本由 AI 还是人类撰写、由哪家实验室的模型撰写,以及由哪个人撰写。
Claude Fable 5.2 在 398 篶作者身份从未公开的非公开同行评审意见中,识别出了 56% 的评审人。对于 1 年内发帖超过 500 次的 Reddit 用户所写的评论,它识别出了其中 69% 的评论背后的账号。
如果你的产品依赖匿名写作,例如盲审或化名论坛,就应认为作者的匿名性较弱,并用前沿模型对其进行测试。
原文
New University of Maryland paper finds that frontier models now detect AI-written prose as well as trained detectors like Pangram, while also explaining their reasoning. Can name a text's author, AI lab or specific person, from that text alone. Anonymous peer review and pseudonymous accounts assume writing alone doesn't reveal who you are. Spotting AI text was also thought to need trained detectors, not plain prompting. The authors gave models a single text and a zero-shot prompt, with no web search, tools, or candidate list. They tested AI-or-human detection, which lab's model wrote it, and which person wrote it. Claude Fable 5.2 named the reviewer of 56% of 398 private peer reviews whose authors were never made public. It named the account behind 69% of comments from Redditors with over 500 posts in a year. If your product depends on anonymous writing, such as blind review or pseudonymous forums, treat author anonymity as weak and test it against frontier models.