NOTAI.AI: Explainable Detection of Machine-Generated Text via Curvature and Feature Attribution
Researchers combine curvature analysis, neural scores, and stylometrics in an explainable detection system with a live web demo.
A team of researchers has introduced NOTAI.AI, a novel and explainable framework designed to detect whether a piece of text was written by a human or an AI model. The system builds upon the Fast-DetectGPT method by integrating a sophisticated blend of 17 distinct, interpretable features. These include Conditional Probability Curvature, a ModernBERT detector score, various readability metrics, and stylometric cues. All these signals are fed into a gradient-boosted tree (XGBoost) meta-classifier, which makes the final determination. This multi-faceted approach aims to create a more robust detector than those relying on a single signal.
Beyond just providing a binary classification, NOTAI.AI's core innovation is its commitment to explainability. It employs Shapley Additive Explanations (SHAP) to calculate both local and global feature-level attributions, showing exactly which factors contributed most to the decision. These technical attributions are then translated into structured, natural-language rationales through a dedicated LLM-based explanation layer, making the system's reasoning accessible to non-experts. The framework is deployed as a fully functional, interactive web application where users can input text for real-time analysis, visually inspect how different features influence the verdict, and review the structured evidence presented. The team has made the source code and a demo video publicly available to support reproducibility and further research.
- Combines 17 interpretable features including curvature, neural scores (ModernBERT), and stylometrics in an XGBoost classifier.
- Uses SHAP for feature attribution and an LLM layer to generate natural-language explanations for its detections.
- Deployed as an interactive web app for real-time analysis, allowing users to inspect the visual evidence behind each decision.
Why It Matters
Provides a transparent, auditable tool for detecting AI-generated content, crucial for academic integrity, content moderation, and trust in digital information.