Beyond Content: How Grammatical Gender Shapes Visual Representation in Text-to-Image Models

TL;DR

GramVis investigates whether grammatical gender shapes visual representations in multilingual text-to-image models, revealing robust male-leaning biases driven by masculine grammatical cues and more variable effects for feminine cues. The approach uses a cross-linguistic dataset of 800 gender-divergent words across seven languages and evaluates three advanced T2I models, generating $28{,}800$ images under controlled prompt templates. The findings show that language resource availability and model architecture systematically modulate these effects, with high-resource languages and Flux-like models exhibiting stronger associations. This work demonstrates that language structure itself meaningfully biases AI-generated visuals, offering a new dimension for assessing and mitigating bias in multilingual multimodal AI.

Abstract

Research on bias in Text-to-Image (T2I) models has primarily focused on demographic representation and stereotypical attributes, overlooking a fundamental question: how does grammatical gender influence visual representation across languages? We introduce a cross-linguistic benchmark examining words where grammatical gender contradicts stereotypical gender associations (e.g., ``une sentinelle'' - grammatically feminine in French but referring to the stereotypically masculine concept ``guard''). Our dataset spans five gendered languages (French, Spanish, German, Italian, Russian) and two gender-neutral control languages (English, Chinese), comprising 800 unique prompts that generated 28,800 images across three state-of-the-art T2I models. Our analysis reveals that grammatical gender dramatically influences image generation: masculine grammatical markers increase male representation to 73% on average (compared to 22% with gender-neutral English), while feminine grammatical markers increase female representation to 38% (compared to 28% in English). These effects vary systematically by language resource availability and model architecture, with high-resource languages showing stronger effects. Our findings establish that language structure itself, not just content, shapes AI-generated visual outputs, introducing a new dimension for understanding bias and fairness in multilingual, multimodal systems.

Figures (11)

• Figure 1: Grammatical gender affects T2I outputs. Top: feminine‐gendered "guard" (une sentinelle / die Wache) yields more feminine imagery than English. Bottom: masculine‐gendered "gossip" (un commérage / der Tratsch) produces more masculine visuals than English, illustrating how language structure influences visual representation.
• Figure 2: Our GramVis benchmark features gender-divergent words across five gendered languages (FR= French, DE= German, ES=Spanish, IT=Italian, RU=Russian). Left: grammatically feminine words represent stereotypically masculine concepts, such as "die Autorität" ("authority") in German. Right: grammatically masculine words represent stereotypically feminine concepts, such as "un commérage" ("gossip") in French.
• Figure 3: GramVis dataset creation pipeline: (1) Word identification - Collecting gender-divergent words across five gendered languages where grammatical gender contradicts stereotypical associations; (2) Expert validation - Linguists and annotators verify gender divergence through dictionaries and human judgment; (3) Prompt engineering - Designing gender-neutral prompts that only inherit grammatical gender from inserted target words; (4) Cross-linguistic image generation - Creating images using identical semantic content expressed in both gendered and gender-neutral languages; (5) Structured analysis - Classifying visual outputs to measure how grammatical gender influences representation.
• Figure 4: Qualitative examples demonstrating expected grammatical gender effects, where feminine grammar increases female representation and masculine grammar increases male representation compared to gender-neutral baseline prompts.
• Figure 5: Qualitative examples demonstrating counterintuitive grammatical gender effects, where feminine grammar decreases female representation and masculine grammar decreases male representation compared to gender-neutral baseline prompts.