Pangram Labs released Pangram 4, an AI-text detection model that claims state-of-the-art performance on the task of identifying machine-generated content. The numbers they're citing are strong: an AUROC of 0.9916, which means the model separates AI-written from human-written text with high accuracy. The false positive rate sits at 0.0041% (flagging human text as AI), while false negatives come in at 0.3396% (missing actual AI text). These metrics represent a meaningful jump from Pangram 3, their previous version. What makes Pangram 4 different goes beyond raw accuracy numbers. The model handles out-of-distribution cases better—meaning it doesn't fall apart when tested on text styles or domains it wasn't trained on. It also resists adversarial attacks, which matters because people actively try to fool detection systems by tweaking their prompts or using specific writing tricks. More practically, Pangram 4 can spot fine-grained edits and mixed content where humans and AI worked together on the same piece, which is closer to how people actually use these tools in the real world. The report doesn't provide much detail about methodology, training data, or how they tested these claims. It reads like a technical announcement rather than a full research paper, so there's no way to assess whether these results hold up across independent testing or what tradeoffs they made to achieve these numbers. The emphasis on detecting interleaved human-AI collaboration is worth watching, since that's where detection gets genuinely hard.