How Pangram Checks Writing — A Simple Guide for Learners

Pangram has become a widely used statistical detector that assigns each sentence a score for being human-written, AI-assisted, or AI-generated, and its line-by-line reports have been adopted by platforms and conferences. The system's designers have published technical notes and a Pangram 4 report on arXiv dated July 29, 2026, which details model architecture and benchmark performance. While Pangram claims an industry-leading low false positive rate, independent researchers have both praised its accuracy on clear-cut AI outputs and warned about performance under adversarial editing.

In a June 29, 2026 study in the International Journal for Educational Integrity, Pangram was compared with GPTZero, Turnitin and Copyleaks and was found to align closely with ground-truth labels in many controlled tests. However, an arXiv paper on test-time adaptation reported that Pangram detected only 24.1% of adversarially modified AI text in their experiments, highlighting how distribution shifts can reduce detection rates. Consequently, institutions that use Pangram for screening have been advised to combine automated flags with human review.

Pangram has released an image-detection research preview and a separate blog post discussed watermarking and image provenance; the image model showed strong benchmark AUROC scores on some datasets but also made real-world mistakes when AI images appeared inside photographs. Substack integrated Pangram on July 21, 2026, enabling readers to scan posts and comments for AI content, which spurred debate about reliance on automated labels.

Ultimately, Pangram is useful as a technical tool to flag likely AI involvement, but it cannot prove intent or full authorship; its outputs should be interpreted as probabilistic signals that require context, corroboration, and care before being used in high-stakes decisions.