Fast-AI-detector: a fast local CLI for detecting AI-generated text (github.com)

The newly launched Fast-AI-detector is a command-line interface (CLI) tool designed to quickly assess whether text was generated by an AI. It employs a compact distilled transformer model, which has approximately 40 million parameters, to mimic the insights of a much larger 4 billion parameter Gemma model. This tool offers two main operational modes: a contrast-based prediction system and a RAID-finetuned classifier, with the contrast mode recommended for typical use due to its better calibration around the human/AI distinction. The significance of Fast-AI-detector lies in its ability to leverage a lightweight model without sacrificing performance, making it accessible for local deployment. Testing on the RAID validation dataset has shown high accuracy with the RAID-finetuned model achieving 96.42% balanced accuracy and a ROC-AUC score of 0.9958. Key features include a user-friendly output system that highlights the human/AI classification along with supporting SAE (Structure-Aware Embeddings) annotations. This innovation not only enables fast detection of AI-generated content but also retains interpretability elements, crucial for understanding model decisions and enhancing trust in AI systems.