Visual Sanad: Auditing Vision-Language Models for False Acceptance of Qur'anic Citations
This study tested how well various vision-language models handle verifying if a displayed religious quote actually matches its cited reference from the Qur'an. It found that while top models perform well, many open-weight models struggle with weak discrimination, often defaulting to 'VALID' even when citations are incorrect. The core finding is a failure in culturally grounded multimodal factuality where fluent models treat religious citations as self-authenticating without strong verification.