EVENT DOSSIER
NotAI.AI is an interpretable system designed to detect machine-generated text, with an F1 score of 0.9685 on the test set. The system combines sentence-level conditional probability curvature, neural detector ratings, and interpretable style and readability features. It uses XGBoost meta-classifiers for prediction and TreeSHAP to calculate feature contributions, generating natural language explanations. The research team evaluated the model on a subset of RAID datasets including human-written, cleanly generated by AI, and attacked AI-generated texts. Results showed that the model outperformed variants based on a single feature family. Additionally, automatic evaluations indicated that 94.5% to 98.6% of the explanations agreed with detector evidence, according to the two model judges. The web interface, source code, and demonstration videos of NotAI.AI are now available publicly.
NotAI.AI NotAI.AI 发布可解释机器生成文本检测系统
NotAI.AI 是一款可解释的机器生成文本检测系统，其 F1 分数在测试集上达到 0.9685。该系统结合句子级条件概率曲率、神经检测器评分以及可解释的风格和可读性特征，利用 XGBoost 元分类器进行预测，并通过 TreeSHAP 计算特征贡献以生成自然语言解释。研究团队在包含人类写作、干净 AI 生成及受攻击 AI 生成的 RAID 数据集子集上进行了评估，结果显示该模型优于基于单一特征家族的变体。此外，自动评估表明两个模型法官认为 94.5% 至 98.6% 的解释忠实于检测器证据。NotAI.AI 的网页界面、源代码及演示视频已公开可用。