New unified AI detection framework uses Mahalanobis distance scores for LLM text and hallucinations

A new paper from arXiv (2607.07527) by Zhang, Hu, Peng, and Tian introduces a unified detection framework for AI-related content and artifacts. The method leverages Mahalanobis distance scores (MDS) to characterize the 'positive class' (e.g., human-generated text, factual statements, unwatermarked text, non-adversarial samples). The core challenge is accurately estimating the covariance matrix of deep representations of positive samples, which are often multi-class and exhibit both homogeneity and heterogeneity. The authors develop joint estimation methods for both casewise and cellwise minimum covariance determinant (MCD) estimators, providing efficient optimization algorithms and proving convergence. They also define a breakdown point for these joint estimators and prove high breakdown robustness.

The framework is designed to work across four critical AI detection tasks: detecting LLM-generated text, identifying hallucinations, spotting watermarks, and recognizing adversarial examples. This unified approach is significant because it replaces ad-hoc detection methods with a principled statistical tool that handles diverse data structures. The authors provide theoretical guarantees and empirical validation, showing that the method performs competitively across all four settings. For practitioners, this means a single, mathematically rigorous system could replace multiple separate detectors, simplifying deployment and improving reliability in AI governance and security applications.