Technical Report on the Pangram AI-Generated Text Classifier (2024) (arxiv.org)

A new technical report introduces Pangram Text, a transformer-based neural network designed to identify AI-generated text versus human-written content. This classifier has demonstrated a remarkable 38 times lower error rate compared to leading commercial detection tools and zero-shot methods like DetectGPT across a diverse benchmark of 10 text domains, including academic and creative writing. The innovation lies in a novel training algorithm called hard negative mining with synthetic mirrors, which significantly reduces false positives, particularly in data-rich domains such as product reviews. The significance of Pangram Text for the AI/ML community is profound, as it not only enhances the accuracy of text classification but also ensures fairness by not displaying bias against nonnative English speakers. Its ability to generalize effectively to unseen domains and models highlights the potential for broad applications in content moderation, academic integrity, and combating misinformation. This advancement sets a new standard for classification tools in an era where distinguishing between AI and human-generated content is increasingly critical.