Pangram 4.1 Model Card (www.pangram.com)

Pangram Labs has announced the release of Pangram 4.1, an enhanced version of its AI detection model, on October 19, 2026. This latest model boasts significant improvements in accuracy and performance under uncertain conditions compared to its predecessor, Pangram 4. Among the key updates is the addition of a new output feature, class_probabilities, which provides insights into the model's confidence in its label predictions. Pangram 4.1 maintains the token-level prediction structure from earlier versions, categorizing text segments as Human, AI-Assisted, or AI-Generated, while also demonstrating a slightly improved resolution for detecting shorter text segments. The advancements in Pangram 4.1 are particularly relevant for the AI/ML community, as it addresses challenges that detection models face, such as false positive rates, which have been sharply reduced across multiple languages. The model's accuracy is backed by robust evaluations, with false positive rates dropping to 0.0090% for English and 0.0033% for a comprehensive multilingual dataset. Furthermore, the model shows greater stability in predictions, especially under minor text edits, demonstrating a significant decrease in unintended verdict flips. These enhancements underscore Pangram Labs' ongoing commitment to improving the reliability of AI detection while acknowledging the ethical implications of false accusations stemming from AI usage evaluations.