Pangram: similarity does not constitute proof
Pangram is what is known as a classification model. Its training involves comparing human-written texts with output from generative models. The aim is to learn the statistical characteristics that enable them to be distinguished. To achieve this, the company uses, amongst other things, a technique that involves creating ‘syntheticmirrors ’. These are texts produced by AI to match the content of human examples. This… comparison aims to avoid a crude oversimplification, such as associating a particular subject with an artificial origin. The AI therefore reads texts written by humans and AI versions created from those texts. Its method also includes a correction loop known as ‘hard negative mining ’. The detector examines a pool of human texts; those it incorrectly flags are retrieved, accompanied by new artificial equivalents, and then…