Detecting Claude by counting letters

This blog post discusses the development of a text detection system aimed at identifying content generated by the Claude Opus 5 AI model. The author details a technique based on n-grams, where character sequences of length 4 are counted to differentiate human writing from AI-generated text. Data for training the model was gathered from pre-2021 sources, and the system shows a high accuracy rate in distinguishing between human and AI text. The author shares insights from their background in questioned document examination and the efficiency of the n-gram approach in achieving reliable detection.