Attackers have been abusing style transfer techniques with Large Language Models, significantly deteriorating the performance of their fake news detection systems. Most of the few existing methods have concentrated on mitigating this issue using output alignment. We argue, however, that this paradigm falls short of addressing issues arising in the latent representation space. To go beyond these limitations, we start developing the Content-Style Invariant Detector (CSID): a new paradigm aiming at realigning efforts from robust prediction to representation-level semantic consistency. It introduces this paradigm shift by two novel components: the Semantic Anchoring method, which clusters style variations into a multi-view supervised setting, along with Contrastive Representation Stabilization—a contrastive model, specifically crafted to align content-similar variations to converge into a compact style-independent grouping in vector space. Experiments on PolitiFact, GossipCop, and the LUN Dataset establish CSID as a new performance bar concerning stability. Moreover, CSID shows no performance degradation—less than 1% loss in Accuracy on the adversarial GossipCop Restyle dataset, apart from improving test accuracy on the authentic test sets, outperforming state-of-the-art competitors and thus validating the effectiveness of emphasizing representation-based theory over output alignment methods.