<p>The rapid evolution of image generation techniques has benefited several fields, but it has also given rise to security concerns. As a countermeasure, numerous AI-generated image detectors have been developed successfully. However, existing methods exhibit limited generalization and robustness in handling the continual emergence of new generative models, owing to the inability to capture the intrinsic representational discrepancy of synthetic content. To address these challenges, a novel discrepancy-driven detection framework called Discrepancy-Aware Dual-Path Network (DADP) is proposed for AI-generated image detection. Specifically, the unstable discrepancy between real and synthetic images under controlled perturbations is first formulated as a learnable signal, which is used to modulate attention and feature fusion. Then, a Discrepancy-Aware Attention Mechanism (DA-CBAM) is meticulously engineered for a dual-branch frequency–semantic feature extractor, where the discrepancy signal adjusts channel and spatial attention to localize perturbation-sensitive artifact regions. Finally, a Discrepancy-Guided Evidence Aggregation module is devised to compute discrepancy-based weights and adaptively fuse multi-level frequency and semantic features. Consequently, the proposed DADP framework simultaneously advances the generalization and robustness of AI-generated image detection. Extensive experiments demonstrate that DADP sets a new state-of-the-art, achieving a 1.94% improvement in cross-generator generalization accuracy while exhibiting superior robustness against various image degradations.</p>