When Detectors Forget Forensics: Blocking Semantic Shortcuts for Generalizable AI-Generated Image Detection

Researchers have proposed new approaches to improve the detection of AI-generated images, addressing the challenge of distinguishing between real and synthetic content.

The growing realism of generative models has made it increasingly difficult to detect AI-generated images, with large-scale pre-trained Vision Foundation Models struggling to generalize to unseen generation pipelines^[1]. A key failure mechanism, termed 'semantic fallback,' occurs when forensic fine-tuning fails to reshape the representation space. To address this, researchers have proposed a Geometric Semantic Decoupling (GSD) framework, which suppresses semantically dominant directions to promote invariant forensic representations^[1]. GSD leverages a frozen CLIP encoder and Singular Value Decomposition (SVD) to estimate and suppress semantic components. Another approach, called DRIFT, measures robustness gaps in frozen vision foundation models by decomposing representation space into complementary robust and fragile subspaces^[2]. DRIFT outperforms training-free robustness-based baselines in open-world generalization and provides interpretable invariance-violation maps. The research was submitted in 2026, with one paper revised on 3 Jun 2026, achieving a 15 times reduction in computational overhead. The submissions were made on arxiv.org, with the initial v1 submission being 15,462 KB in size and the v2 submission being 8,021 KB^[1].