Artificial intelligence now plays a kind of key role in graphic design, marketing, and everyday social platforms where images produced from a line of text appear almost indistinguishable from normal photos. That convenience, though, comes with consequences that are not very visible unless someone examines the output closely.
A new multilingual study from researchers in Germany and partner institutions reveals that text prompts written in different languages can influence the gender presentation of generated faces, and these shifts are not random at all. The underlying systems amplify familiar stereotypes in occupations and personality traits, turning assumptions into visual results. The investigation shows that no matter how advanced modern text to image generators have become, they still reflect and sometimes intensify cultural patterns about gender roles.
Testing Nine Languages and Thousands of Prompts
To understand how language structures interact with model behavior, the research team developed a benchmark that compares outputs across languages with distinct grammatical design.
The benchmark is known as the Multilingual Assessment of Gender Bias in Image Generation. It evaluates occupations and descriptive adjectives with carefully controlled phrasing. The set includes languages that mark gender directly in nouns such as German, Spanish, French, Italian, and Arabic. It also includes English and Japanese which primarily carry gender through pronouns rather than the form of the occupation word. Korean and Chinese are present as well, representing languages without grammatical gender in nouns or pronouns. This wide linguistic range allowed the researchers to investigate whether the same job title or description leads to similar images when prompts are identical in content.
Prompt Structure Can Influence Visual Interpretation
The benchmark uses different prompt types to observe how small language choices affect results.
One type refers to an occupation using the default noun that traditionally acts as a generic masculine term in languages that rely on grammatical gender.
Another type avoids the occupation noun entirely by replacing it with a description of the work that a person performs.
Feminine versions of job titles appear in languages where they exist. In German, there is even a gender star notation that tries to make references more inclusive by altering the written form of a word with a special character. These choices were introduced to learn whether changing prompt structure reduces bias or whether the models continue showing strong patterns even when language attempts to remove gender cues.
A Large-Scale Image Evaluation Process
The study tested five multilingual image generation models that are widely known for high resolution output and sophisticated language understanding. All systems were given 100 attempts for each text prompt and produced images intended to show identifiable human faces. With more than 3600 prompt variations and a hundred generated samples each time, over 1.8 million images were analyzed. The model outputs were then classified to determine the perceived gender in every portrait.
Researchers measured how far the results deviated from an equal presentation of male and female appearances. A measure of absolute deviation from balance helped indicate how strongly stereotypes emerge when the model interprets a role like accountant, nurse, firefighter, or software engineer.
Bias Patterns Show Up Consistently Across Models
The outcomes confirm that gender distribution in generated images rarely matches a balanced expectation, and the strength of the skew varies by language. For jobs viewed as masculine in many societies, such as engineering or accounting, most images portrayed male presenting individuals even when the text did not indicate gender. Jobs associated with caregiving or service often shifted the distribution strongly toward female presenting individuals.
These tendencies appear repeatedly across different platforms tested, which suggests that the bias comes from common exposure to large datasets shaped by real world social structures. The study found that some languages produced noticeably stronger stereotypes than others, yet the level of grammatical gender in the language did not reliably predict the degree of bias. Shifting from one European language to another could change the portrayal significantly even when both languages handle gender in similar ways.
Gender Neutral Phrasing Reduces Bias but Creates New Challenges
Prompts that avoid gendered nouns sometimes reduce the size of the imbalance, although the improvement is not enough to reach fairness. When occupations are rewritten so that the prompt describes the work without using a direct title, the model can lose some clarity and create images with more background scenes and fewer clear facial features. That shift affects how well the prompt and the image correspond in meaning. Systems also needed more attempts to produce a recognizable face from these longer and more complex prompts. As a result, choosing neutral style text becomes a tradeoff. The output may contain less amplified gender bias, yet the purpose of the request may not always be met if someone expects stability and accuracy in the final image.
Language Choices That Try to Ensure Fairness May Backfire
Methods introduced by human language communities to make job titles more inclusive do not always help when used in AI prompts. In the case of the German gender star approach, the models produced even more female appearing faces in several occupations rather than a balanced set. This suggests that inclusive writing styles might be underrepresented in training data, causing the model to rely on the parts of the word that it recognizes more strongly, which can shift interpretation rather than neutralize it.
More Attention Needed for Global Fairness
The researchers emphasize that users outside the primary training language may encounter biased performance precisely because their prompts are in languages that the model does not interpret as reliably. Out of distribution languages sometimes produced images that barely matched the job description at all, which can lower the measured bias only because meaningful gender cues are missing. With generative systems becoming accessible throughout regions with diverse language traditions, fairness concerns must go beyond English centric design.
Bias Remains a Persistent Issue in Image Generation
This multilingual evidence highlights the limits of simple prompt rewriting as a solution to gender imbalance. Even when prompts attempt to conceal gendered cues, representation patterns stay uneven. The findings call for stronger tools and deeper attention to training choices in text to image models, because language alone cannot remove stereotypes already ingrained in data. A globally deployed generation system that portrays individuals in occupations ought to provide imagery that does not reinforce narrow assumptions linked to gender. The results show how crucial it will be to improve both multilingual understanding and fairness control as the technology becomes a standard part of communication and creativity.
Notes: This post was edited/created using GenAI tools.
Read next: Wikipedia Faces Political Pressure As Co-founder Renews Bias Claims
