NOTAI.AI: Explainable Detection of Machine-Generated Text via Curvature and Feature Attribution

Explainability matters for AI detection: Rather than simply labeling text as "AI-generated" or "human-written," NotAI.AI shows users which specific features influenced each prediction, making the detection process transparent and trustworthy rather than a "black box" verdict.

Multiple detection signals are more reliable: The system combines three different approaches—probability patterns in sentences, neural network analysis, and writing style features—achieving 96.85% accuracy (F1 score), demonstrating that layering different detection methods produces stronger results than relying on any single approach.

AI-generated explanations can be faithful to evidence: The system automatically converts technical detection evidence into plain-language explanations that human judges rated as accurate 94.5-98.6% of the time, suggesting AI systems can communicate their reasoning in understandable ways when properly designed.

Adversarial attacks on AI detectors are a real concern: The evaluation specifically tested "attacked AI-generated texts"—content deliberately modified to evade detection—highlighting that as AI generation improves, detection systems must anticipate and defend against intentional circumvention attempts.

Responsible AI detection requires public transparency: By releasing the code, web interface, and demonstration video publicly, the researchers enable independent verification and broader adoption, supporting the principle that AI safety tools should be openly scrutinized rather than proprietary black boxes.