Vision-language models are being deployed in settings where accuracy carries cultural weight, including religious and educational contexts. Yet their ability to verify whether a cited text actually appears in an image remains largely untested, according to a new arXiv preprint.

The authors introduce Visual Sanad, an auditing approach designed to probe models for false acceptance of Qur'anic citations. The work highlights a specific failure mode: a model may confidently accept a citation as matching the visual content when it does not, which could mislead users in sensitive applications.

The abstract does not yet report experimental results, so the paper's current contribution is the problem formulation and the proposed audit method. Its significance lies in drawing attention to a high-stakes domain where verification errors could have real-world consequences.