AI Vision Systems Overlook Billions of Multiscript Language Speakers
Researchers created a new test showing AI vision systems struggle with languages written in multiple scripts. The study highlights how current AI models unfairly disadvantage billions of people who use different writing systems for the same language. However, the original source does not support testing this with tools like Google Lens or Microsoft Seeing AI, as those are not explicitly mentioned.
Researchers at ArXiv cs.CL released a new benchmark called PuMVR (Punjabi Multimodal Visual Reasoning) to test AI vision systems. These systems, which combine image recognition with language understanding, often assume one language equals one writing system. The benchmark shows these models fail to handle languages like Punjabi, Serbian, and Hindi-Urdu, which use multiple scripts.
This matters because billions of people worldwide use languages that can be written in different scripts. For example, Punjabi can be written in both Gurmukhi and Shahmukhi scripts, while Hindi-Urdu uses Devanagari and Perso-Arabic scripts. Current AI systems can't properly understand or process these variations, putting these users at a disadvantage. The benchmark consists of 375 culturally grounded image-reasoning questions designed to quantify this orthographic bias.