Pangram 4 Technical Report [pdf] (pangram-public.s3.us-east-1.amazonaws.com)

Pangram Labs has unveiled Pangram 4, a state-of-the-art AI text classification model that achieves remarkable performance metrics, including an AUROC of 0.9916, a false positive rate of just 0.0041%, and a false negative rate of 0.3396%. This latest iteration surpasses its predecessor, Pangram 3, in accuracy and shows enhanced generalization capabilities, specifically in identifying AI-generated content amid a deluge of online text. Pangram 4 is particularly adept at recognizing subtle AI-human co-authored texts and handling adversarial attacks, making it a significant advancement in the AI/ML landscape. The emergence of Pangram 4 is crucial as the volume of AI-generated text continues to escalate, complicating the task of discerning genuine human writing from AI output. With Pangram 4's improved boundary detection and performance on mixed authorship text, it addresses the pressing need for dependable AI monitoring systems to mitigate the societal harms of AI-generated misinformation. Furthermore, it employs a sophisticated training methodology using a diverse dataset, which includes human and AI-generated text across various domains, reinforcing its robustness and applicability across different writing styles and contexts.