The Quiet Colonization of Prompt Generation

Bias in artificial intelligence is often discussed in the context of explicit stereotypes—obvious cases where a system produces harmful or offensive outputs. But sometimes, bias shows up in quieter, subtler ways. During a recent experiment, I found an example that underscores how cultural defaults can persist even when a prompt seems perfectly neutral.

The Experiment

I started with a simple idea: create image prompts that described everyday, culture-agnostic scenes—moments of quiet gathering, evening walks, open landscapes. Nothing in these descriptions referenced a particular region, ethnicity, or cultural context. To reduce any hint of “English-language bias,” I translated each prompt into Kiswahili, a major African language spoken across East Africa.

To make the instruction clear to the model, every prompt began with the phrase:

“Tengeneza picha ifuatayo” — “Generate the following image.”

I provided only the Kiswahili text after that opening line, never mixing in English or specifying a location. Here are several examples of the exact prompts I used, shown in both English and Kiswahili for reference:

English Prompt	Kiswahili Prompt
“A quiet dinner gathering with a small group of people. Capture only the feeling of togetherness and the sense of time passing.”	“Tengeneza picha ifuatayo: Chakula cha jioni tulivu kinachoshirikisha kikundi kidogo cha watu. Shika tu hali ya pamoja na hisia za wakati kupita.”
“A serene outdoor moment on the edge of a wide landscape. Focus on the sense of space, light, and a soft breeze without naming a place, weather, or time of day.”	“Tengeneza picha ifuatayo: Wakati wa utulivu nje, pembezoni mwa mandhari pana. Lenga hisia za nafasi, mwanga, na upepo laini, bila kutaja eneo maalum, hali ya hewa, au muda wa siku.”
“A lone traveler walking along a quiet path as evening falls. Only the sound of footsteps and a distant wind define the journey.”	“Tengeneza picha ifuatayo: Msafiri peke yake anatembea kwenye njia tulivu wakati jioni inashuka. Safari inaendelea kwa utulivu, ikitambulika tu kwa sauti ya hatua na upepo wa mbali.”

Despite the neutral wording and the use of Kiswahili, the image outputs consistently skewed toward Western aesthetics: architecture resembling small American towns, individuals dressed in contemporary Western fashion, and lighting reminiscent of U.S. or European photography styles.

Observing the Pattern

This pattern reflects a well-known phenomenon in AI: models inherit the dominant cultural signals of their training sets. Large image datasets and caption corpora are heavily weighted toward Western, especially American, sources—stock photography, popular social media, and English-language content.

Even when a user provides a neutral or foreign-language description, the model’s “mental map” of the world gravitates to what it has seen most. The result is a kind of “silent default,” where Western imagery emerges as the baseline for everyday human life.

It’s not that the system ignores other cultures on purpose. Rather, it has learned that “default human life” looks a lot like what it encountered most frequently during training. Because English remains the lingua franca of the internet and many large datasets are curated in the United States, those defaults inevitably lean Western.

Company Origin and Broader Context

ChatGPT is developed by OpenAI, a U.S.-based company headquartered in San Francisco. That fact alone doesn’t dictate the model’s outputs, but it reinforces the reality that much of the model’s development and evaluation happens within an American cultural frame. Combined with the prevalence of U.S.-sourced images on the open web, this creates a feedback loop: the data available for training reflects the culture that produces and hosts the most content, which in turn reinforces the same cultural defaults in the model.

Next Steps

I plan to expand this experiment to other AI image engines to see whether they reveal the same tendencies. Using the exact same Kiswahili-only prompts—each beginning with “Tengeneza picha ifuatayo”—I will test Midjourney, Adobe Firefly, and other emerging platforms.

If similar Western-centric imagery continues to surface, it will highlight not only the need for more geographically balanced training data but also the importance of transparency about cultural defaults in AI systems.

Understanding these subtle biases is essential if we want AI to truly represent the world’s diversity—not just the culture most overrepresented in its data.