A word written in cursive can change how an AI defines that word compared to the same word in a clean font.
Multimodal models are expected to separate the semantic meaning of text from its aesthetic presentation. This research proves that the visual style of text leaks into the model attribute-based descriptions. A gritty font might make the AI describe a concept as more aggressive or dark even if the word itself is neutral. This failure of abstraction means AI does not perceive words as pure data like humans do. It suggests that visual branding and typography are powerful, unintended levers for controlling AI perception. Developers must account for these aesthetic leaks when building vision-language systems.
Revealing the Impact of Visual Text Style on Attribute-based Descriptions Produced by Large Visual Language Models
arXiv · 2604.27553
When the visual style of text is considered, a wide variety can be observed in font, color, and size. However, when a word is read, its meaning is independent of the style in which it has been written or rendered. In this paper, we investigate whether, and how, the style in which a word is visualized in an image impacts the description that a Large Visual Language Model (LVLM) provides for the concept to which that word refers. Specifically, we investigate how functional text styles (readability