Detecting the Machine: Comprehensive Benchmark of AI-Generated Text Detectors Across Architectures, Domains, and Adversarial Conditions

arXiv 2603.17522·medium signalXBlueskyLinkedInCopy link

Large-scale evaluation of AI-generated text detectors across diverse LLM architectures, writing domains, and adversarial evasion attacks — addressing the critical gap between lab detector performance and real-world deployment reliability. Finds systematic detector failures under domain shift and paraphrasing attacks, with no single detector generalizing across all conditions. Provides practitioners with a decision framework for selecting detectors based on threat model rather than aggregate accuracy.