Pangram Has Emerged as the Gold Standard of AI Detection. Should You Trust It?
As our interview progresses, Spero’s responses seem sticky, stopping and starting, and not just because he’s eating. I ask about his hiring ethos. Spero murmurs “hmm” before turning away without apology to microwave his food. Fifteen long seconds pass in silence. He finally faces me again and says, “The average person is at Pangram because they care about the mission.” To detect AI, Pangram uses a method called “synthetic mirroring” by which it takes human writing and has LLMs generate a close match. This teaches its model how AI writes. Pangram also uses “hard negative mining,” searching datasets for false positives that it can synthetically mirror and use to augment its training set—using mistakes to retrain the machine. “All of our datasets are properly licensed, which I think is kind of rare in the AI world today,” Spero says, though that’s partly because Pangram’s product is far less data-hungry than ChatGPT or Claude. “I don’t want to completely throw the AI companies under the bus,” he adds, “but I think they’ve lost a lot of …
