一篇文章，到底多少是 AI 改的？
Wofo Ink 逐句判定 AI 改写的准确率，显著高于 GPTZero 和 Pangram。
测试范围：100 篇被 AI 部分改写的文章，共 2,273 句。
Bridges across the nation are aligned with the roads they carry, which means that some have a curved horizontal layout. The analysis and design of these bridges is, therefore, a new design activity for bridge engineers and, in addition, very little experimental data are available to substantiate the design with respect to the dynamic response of these bridges during an earthquake. This research will be conducted alongside an ongoing project collecting field measurements of the dynamic behavior of a curved box bridge in New Zealand. Researchers will measure the bridge’s distortion and warping modes and add to existing findings on soil–structure interaction and the coupling between the superstructure’s flexural and torsional dynamics. The resulting measurements will support the creation of analysis techniques and computer software, as well as provide a foundation for design standards for highway bridges.
逐句对照
同一篇文章，四种读法。
下面每一篇都是人写的原文被 AI 改写了一部分。第一行是真实答案，后三行是各家检测器逐句标出的结果。
Bridges across the nation are aligned with the roads they carry, which means that some have a curved horizontal layout. The analysis and design of these bridges is, therefore, a new design activity for bridge engineers and, in addition, very little experimental data are available to substantiate the design with respect to the dynamic response of these bridges during an earthquake. This research will be conducted alongside an ongoing project collecting field measurements of the dynamic behavior of a curved box bridge in New Zealand. Researchers will measure the bridge’s distortion and warping modes and add to existing findings on soil–structure interaction and the coupling between the superstructure’s flexural and torsional dynamics. The resulting measurements will support the creation of analysis techniques and computer software, as well as provide a foundation for design standards for highway bridges.
100 篇改写稿的全部句子
共 2,273 句：AI 改过的 1,141 句，保持人写原样的 1,132 句。
Pangram 按段落给标签，这一行显示的是每句所在段落的标签。以上是挑选出的例子，完整的 300 篇数据见对比测试。原文出处：NSF 基金摘要（公版）、Enron 邮件公开语料、Stack Exchange（CC BY-SA）。
统计口径：改写稿里，在人写原文中没有原样出现的句子算作 AI 改过。Wofo Ink 以 50% 为界；GPTZero 按它自己的逐句高亮；Pangram 只给段落标签，按句子所在段落计算。在 100 篇纯人写和 100 篇纯 AI 的文章上，三家逐句都几乎全对。
对比测试
在检测器分歧最大的地方做测试。
我们取 100 篇被 AI 部分改写的文章，把各家报出的 AI 比例和真实改动比例相比。300 篇测试集和全部原始分数公开。
离真实 AI 比例的中位差（分），越低越好。 在纯人写和纯 AI 的文章上，三家表现相当。 阅读完整对比并下载数据
批量检测
一次上传二十篇，一张表看完。
把一整个文件夹拖进来，每篇的分数直接列在同一张表里，旁边附逐句分布。想看细节时再点开那一行。