“you might want to check out the leading model Technical Report on the Pangram AI-Generated Text Classifier”
Technical Report on the Pangram AI-Generated Text Classifier
基于 Transformer 的 AI 文本检测器及其低误报训练法
论文介绍 Pangram Text,一个基于 Transformer、用来区分大语言模型生成文本与人类写作文本的神经网络。在覆盖 10 个文本领域、8 个开源和闭源 LLM 的基准上,其错误率比 DetectGPT 等零样本方法和领先的商业检测工具低 38 倍以上;作者提出「合成镜像难负例挖掘」训练算法,使分类器在评论等高数据领域的误报率低几个数量级,并显示它对非英语母语者没有偏见、能泛化到训练中未见的领域和模型。
评论者把它当作较可靠的 AI 文本检测器的出处,引用其报告的极低误报率(如万分之一)和用成对的人类与 LLM 文本训练的做法。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“Our white paper with some methodology is here: Technical Report on the Pangram AI-Generated Text Classifier”
“which is one of the most reliable AI-generated text classifiers out there [2] (with a benchmarked accuracy of 99.85% over long-form text)”
“AI detectors in general are unreliable, but there are a few made by serious researchers that have only 1-in-10000 false positive rate, e.g. Technical Report on the Pangram AI-Generated Text Classifier”
“which strikes me as a good way to approach the problem.”
“they have a research paper here: Technical Report on the Pangram AI-Generated Text Classifier”
看全部 7 条 →
“Pangram is highly effective and has a very low false positive rate.”
- Technical Report on the Checkfor.ai AI-Generated Text Classifier ↗PaulHoule · 2024-03-01 · 1 分
- Technical Report on the Pangram AI-Generated Text Classifier (2024) ↗imustachyou · 2026-07-06 · 1 分