Pangram 4 Technical Report

2026-07-29

AI Summary

This work addresses critical challenges in detecting AI-generated text, including poor out-of-distribution generalization, limited adversarial robustness, and difficulties in identifying human-AI hybrid content. To tackle these issues, the authors propose Pangram 4, a deep learning model that integrates advanced text classification with boundary detection techniques. This approach substantially enhances the model’s ability to discriminate fine-grained edits and interleaved human-AI collaborative texts. Evaluated on standard benchmarks, Pangram 4 achieves an AUROC of 0.9916, with a remarkably low false positive rate of 0.0041% and a false negative rate of 0.3396%. The model demonstrates state-of-the-art performance under both out-of-distribution settings and adversarial attack scenarios, establishing a new benchmark for robust AI-generated text detection.

Abstract

We present Pangram 4, the latest deep-learning-based AI-text classification model from Pangram Labs. We achieve an AUROC of 0.9916 with a false positive rate of 0.0041% and a false negative rate of 0.3396%. In addition to its increased overall accuracy compared with Pangram 3, Pangram 4 exhibits superior out-of-distribution generalization and robustness to adversarial attacks. Another novel contribution of Pangram 4 is its improved ability to distinguish fine-grained edits and mixed AI-human co-authored text. We demonstrate improvements to both boundary detection tasks and the detection of interleaved AI assistance. Finally, we report metrics on standard AI detection benchmarks showing that Pangram 4 achieves state-of-the-art performance on the AI text detection task across a wide variety of settings and domains.

Problem

• AI-text detection
• out-of-distribution generalization
• adversarial robustness
• mixed AI-human text
• fine-grained editing detection

Innovation

• out-of-distribution generalization
• adversarial robustness
• fine-grained edit detection
• AI-human co-authored text
• AI text detection