Text-to-image models globally default to male, enforce gender stereotypes, and penalize female and nonbinary representations
This study presents the largest global analysis of text-to-image models, revealing that they systematically default to male representations, enforce rigid gender stereotypes, and marginalize nonbinary identities, thereby amplifying structural gender biases and causing psychological harm to underrepresented groups.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the quiet hum of a server farm, a new kind of artist is at work. It does not hold a brush or a camera, but instead listens to words and paints pictures. This technology, known as text-to-image generation, has exploded in popularity, creating billions of images in just a few years. These systems are trained on vast collections of photos and text from the internet, learning to connect words like "doctor" or "mother" with the visual details they have seen before. Because these tools are now used by millions of people to create advertisements, stories, and designs, the pictures they produce have become a powerful force in shaping how we see the world. If a machine consistently draws a doctor as a man and a nurse as a woman, it does not just reflect a mistake; it reinforces a belief that these roles belong to specific genders. The question researchers are now asking is not just whether these machines make errors, but what kind of world they are quietly building for us, one image at a time.
A team of researchers set out to investigate this hidden landscape by asking four leading text-to-image models to generate nearly one million pictures. They did not simply ask for random images; they constructed a massive, global experiment. They asked the models to draw people from twenty-two different countries working in over one hundred jobs and performing various daily activities. Crucially, they often left the gender of the person unspecified, asking only for "an Indian person working as a translator" or "a German person gardening." By analyzing the results, the team discovered a consistent pattern: when the prompt did not specify a gender, the models almost always chose to draw a man. On average, 68.6 percent of the images generated from these neutral requests depicted men, a figure that rose even higher for jobs involving authority or power. This default to maleness was not uniform across the globe; it was most pronounced in countries with higher levels of gender inequality and lower income, suggesting the technology absorbs and amplifies the social biases of the places it is used.
The study also looked closely at the quality of the images produced. When the researchers asked for specific genders, they found that pictures of men were consistently rated as higher quality, more realistic, and better understood by the computer than pictures of women. This gap was widest when women were depicted in jobs traditionally held by men. In these cases, the images of women were often judged as less plausible or less realistic than images of men in the same roles. The models also fell into rigid stereotypes regarding how men and women should look. Images of men were frequently shown as serious, muscular, and older, while images of women were almost always depicted as young, thin, smiling, and attractive. This "bidirectional stereotyping" means the machines are not just making men the default, but are actively constructing two very different, and limiting, visual scripts for each gender.
Perhaps the most striking finding concerned people who do not fit into the binary categories of man or woman. When the researchers asked the models to generate images of nonbinary people, the systems struggled significantly. The images produced were often of poor quality, failed to capture the prompt, or defaulted to showing young men. This suggests that for these models, nonbinary identities are not just underrepresented, but are largely illegible. To understand the human impact of these findings, the researchers invited thousands of people to view images of themselves—people who shared the same nationality and gender as the subjects in the pictures. The feedback was revealing. Many participants described feeling a sense of exclusion or insecurity when they saw their identities reduced to narrow, stereotypical templates. One person noted that seeing a "template" of their identity felt strange, as if the machine had a very specific, limited idea of what a person from their country should look like. However, when the images broke these stereotypes—showing a woman in a high-power role or a nonbinary person in a way that felt authentic—participants described feeling affirmed and empowered.
The researchers emphasize that these results are not the fault of a single model or a single company. They tested four different systems from different developers, including those based in Germany, the United Kingdom, and China, and found the same patterns across all of them. This consistency points to a deeper issue rooted in the shared data and methods used to train these technologies. The study concludes that these machines are not neutral tools; they are encoding a set of gender logics that treat men as the standard human, penalize women for stepping outside traditional roles, and erase nonbinary identities. As these systems become more integrated into our daily lives, the images they generate risk becoming a self-fulfilling prophecy, quietly narrowing our collective imagination of who belongs where and who gets to be seen.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.