AI Image Generators Amplify Gender Biases Differently Depending on Language

AI-powered image generators amplify gender stereotypes to varying degrees depending on the language used in prompts, according to a groundbreaking study published today. The research shows that identical profession descriptions produce markedly different gender distributions in AI-generated images when translated across languages, challenging assumptions about algorithmic fairness, reports Mirage News.

Researchers from the Technical University of Munich and TU Darmstadt analyzed text-to-image models across nine languages, developing the first comprehensive benchmark for measuring gender bias in multilingual AI systems. Their findings demonstrate that linguistic structures significantly influence bias patterns, even when grammatical gender rules appear similar.

Language Shapes AI Bias Patterns

The study introduced MAGBIG (Multilingual Assessment of Gender Bias in Image Generation), analyzing over 1.8 million AI-generated images across languages including German, Spanish, French, English, Japanese, Korean, and Chinese. Researchers tested four prompt types: direct profession terms using generic masculine forms, indirect descriptions, explicitly feminine prompts, and gender-neutral phrasings.

Direct prompts with generic masculine terms showed the strongest biases: professions like “accountant” predominantly generated images of white men, while caregiving roles produced female images. Notably, switching from French to Spanish prompts led to a substantial increase in gender bias, despite both languages using similar grammatical structures to mark professions by gender.

Mitigation Efforts Prove Limited

Gender-neutral language and “gender star” conventions — such as the German “Ärzt*innen” for doctors — achieved only slight bias reduction while degrading image quality and text alignment. The study found that prompt design strategies intended to reduce bias were largely ineffective and sometimes even counterproductive.

“Our results clearly show that linguistic structures have a significant impact on the balance and bias of AI-based image generators,” said Alexander Fraser, professor of data analytics and statistics at TUM’s Heilbronn campus. “Anyone using AI systems must realize that different phrasings can lead to entirely different images and thus amplify or mitigate societal role stereotypes.”

Professor Christian Kersting, co-director of hessian.AI, emphasized broader implications: “AI image generators are not neutral — they reflect our biases in high resolution, and this depends critically on language. Especially in Europe, where many languages intersect, this is a wake-up call: fair AI must be developed with language sensitivity in mind.”