Pangram
An AI text detection product that trains a large classifier on contrastive sample pairs of human-written and AI-generated content. It claims 99.98% accuracy and reports approximately 17 million users.
Timeline (last 90 days)
Pangram is a tool positioned for AI content detection, covering the identification of both text and image content.
Pangram's discussion heat on Hacker News was limited, receiving 7 upvotes and 0 comments.
The AI detection tool Pangram gave Druckenmiller's op-ed a 100% probability of being AI-generated.
The organizers reduced Pangram's AI-generation flagging rate from 42.7% to 12.7% by narrowing the text detection window.
NeurIPS's Position Paper Track used the third-party proprietary AI detector Pangram to desk-reject 18.4% of all submissions—178 papers in total—without human review or an appeals process.
NeurIPS did not publish any demographic calibration data when deploying the Pangram system.
Among the rejected papers, 22 were rejected because their detection scores exceeded 0.5 and the authors had checked the declaration stating they had not used AI.
Independent researchers fed recently published papers by three chairs of that track into the Pangram detector, and the results showed AI-generation probabilities ranging from 24% to 69%.
Pangram's default settings initially flagged 42.7% of submissions as 90%–100% AI-generated.