Drifting Away from Truth: GenAI-Driven News Diversity Challenges LVLM-Based Misinformation Detection

TL;DR

This work identifies GenAI-driven news diversity as a key threat to LVLM-based multimodal misinformation detection, introducing DriftBench to systematically study model-level misperception and evidence-level drift. DriftBench provides 16,000 diversified news instances across eight variant types and includes an evidence retrieval component to simulate real-world fact-checking pipelines. Through three evaluation tasks—diversity robustness, adversarial evidence contamination, and reasoning analysis—the study demonstrates substantial average declines in F1 (about $14.8\%$) and destabilized reasoning across both vanilla LVLMs and specialized MMD models, with open-ended diversification and contaminated evidence causing the strongest drops. The findings underscore the need for drift-resilient, diversity-aware MMD approaches and establish DriftBench as a foundation for advancing trustworthy verification in GenAI-adjacent information landscapes.

Abstract

The proliferation of multimodal misinformation poses growing threats to public discourse and societal trust. While Large Vision-Language Models (LVLMs) have enabled recent progress in multimodal misinformation detection (MMD), the rise of generative AI (GenAI) tools introduces a new challenge: GenAI-driven news diversity, characterized by highly varied and complex content. We show that this diversity induces multi-level drift, comprising (1) model-level misperception drift, where stylistic variations disrupt a model's internal reasoning, and (2) evidence-level drift, where expression diversity degrades the quality or relevance of retrieved external evidence. These drifts significantly degrade the robustness of current LVLM-based MMD systems. To systematically study this problem, we introduce DriftBench, a large-scale benchmark comprising 16,000 news instances across six categories of diversification. We design three evaluation tasks: (1) robustness of truth verification under multi-level drift; (2) susceptibility to adversarial evidence contamination generated by GenAI; and (3) analysis of reasoning consistency across diverse inputs. Experiments with six state-of-the-art LVLM-based detectors show substantial performance drops (average F1 -14.8%) and increasingly unstable reasoning traces, with even more severe failures under adversarial evidence injection. Our findings uncover fundamental vulnerabilities in existing MMD systems and suggest an urgent need for more resilient approaches in the GenAI era.

Figures (11)

• Figure 1: Illustration of multi-level drift induced by GenAI-driven news diversity, comprising (1) model-level misperception drift and (2) evidence-level drift.
• Figure 2: Overview of DriftBench, illustrating the construction of diversified news variants and the three evaluation tasks.
• Figure 3: Precision (%) of representative LVLM-based MMD approaches (①: GPT-4o-mini; ②: Claude-3.7; ③: Qwen-VL; ④: CMIE; ⑤: Sniffer; ⑥: Lemma) under all types of controlled and open-ended news diversity.
• Figure 4: Accuracy degradation of MMD methods under increasing evidence drift severity in the DI_OT setting. Higher drift degrees lead to more pronounced performance drops.
• Figure 5: Reasoning explanations generated by LVLMs under controlled settings (R_OI_OT, R_DI_OT), open ended scenarios (F_FI_FT), and cases involving malicious evidence contamination (F_DI_OT).