Text-to-Image Models and Their Representation of People from Different Nationalities Engaging in Activities

This paper investigates how popular text-to-image (T2I) models, DALL-E 3 and Gemini 3 Pro Preview, depict people from 206 nationalities when prompted to generate images of individuals engaging in common everyday activities. The key finding is that these models frequently depict individuals wearing traditional attire or impractical clothing for the specified activities, with these patterns being statistically significantly associated with specific regions and income groups.

Research Objectives

The main objective of the paper is to study how T2I models represent individuals from various countries when prompted to generate images of people engaging in typical activities. Specifically, the paper focuses on three primary research questions:

** RQ1: When generating images of people from various countries performing common activities, to what extent are they depicted in traditional or impractical attire, and how does this correlate with the country's region or income group?**

** RQ2: When employing CLIP, ALIGN, and GPT-4.1 mini to calculate alignment between images and prompts, to what extent do alignment scores correlate with the presence of traditional or impractical attire in labeled images?**

** RQ3: When examining revised prompts, what elements do they introduce to the final instructions, and how do these elements contribute to the representation patterns identified in RQ1?**

Methodology and Dataset Construction

The study constructs a dataset that includes 206 nationalities across five scenarios: couples cooking at home, groups playing soccer in public parks, a person working on their laptop from home, a group jogging, and a group drinking coffee at a coffeeshop. Images were generated using DALL-E 3 (OpenAI) and Gemini 3 Pro Preview (Google DeepMind). The dataset includes 206 images per model across the five scenarios, totaling 2,060 images. The images are manually labeled to indicate whether the depicted individuals are dressed in traditional or impractical attire.

Analysis of Representational Patterns (RQ1)

When aggregating results across all activities and models, 28.4% of the images depicted individuals wearing traditional attire. Among the subset of images assessed for impractical clothing in the two athletics-related activities, 15.9% depicted individuals wearing attire that was impractical for those activities. The prevalence of traditional attire varied by activity; "the cooking and coffee activities yielding the highest percentages of images labeled as “traditional,” whereas the working activity yielded the lowest percentage, likely because it specified a single individual in the prompt."

Statistical Associations with Region and Income Group

The analysis stratified results by region and income groups to identify systematic patterns. The regional analysis revealed that "the Middle East & North Africa region consistently exhibited the highest percentages across activities and label categories. For the 14 activity–label combinations, Middle East & North Africa (MENA) ranked first in the percentage of labeled images in 10 cases and second in the remaining four cases, behind South Asia. Across income groups, Low Income countries exhibited the highest percentages in most scenarios, often followed by Lower Middle Income countries. When combining all 10 sets of results across both models and five activities, 28.4% of images depicted people dressed in traditional attire and strong associations were observed with both region (χ2 = 386.1, p <.001, Cramér’s V = 0.43) and income group (χ2 = 171.8, p <.001, Cramér’s V = 0.28)."

Evaluation Model Efficacy (RQ2)

The second research question examined the effectiveness of CLIP, ALIGN, and GPT-4.1 mini for evaluating alignment between prompts and generated images. For the standard prompts used to generate the images, an unexpected pattern emerged: Across all three scoring methods, alignment scores were consistently higher for images featuring individuals in traditional clothing compared to those without it. This pattern was also held when aggregating results across all activities for both DALL-E 3 and Gemini 3 Pro Preview. For example, for DALL-E 3, the average scores were 31.31 vs. 32.28 (CLIP), 23.11 vs. 25.36 (ALIGN), and 67.63 vs. 75.70 (GPT-4.1 mini).

Analysis of Revised Prompts (RQ3)

The final research question examined the revised prompts generated by OpenAI for the standard prompts across countries, focusing only on DALL-E 3 images. An initial exploratory analysis processed all revised prompts to identify frequently used words, noting that the revised prompts also introduced words not explicitly present in the original prompts, including “traditional” and “diverse.” Specifically, all revised prompts for each activity were processed to count how often the word “traditional” was added, finding it appeared in "88.

Here are specific, actionable improvements to AI systems derived from the findings of this research:

) 1. Implement a Multi-Stage Bias Detection and Mitigation Pipeline in T2I Generation:

The paper demonstrates that representational biases (traditional attire, impractical clothing) are introduced across the entire pipeline—generation, evaluation, and prompt revision.

• [System Improvement]: Integrate a pre-generation bias check layer using metrics derived from RQ1 (labeling criteria) before final output. This layer should flag prompts or generation seeds likely to produce stereotypical outputs based on known regional/income group correlations (MENA, Sub-Saharan Africa, Low Income).

• [System Improvement]: Develop a Prompt Sanitization module that analyzes the generated revised prompts (RQ3 analysis) for the frequent insertion of culturally essentializing terms like traditional, especially when combined with activity prompts that are contextually inappropriate. If such terms appear frequently without explicit user intent, the system should automatically rewrite them to be more context-appropriate or remove them entirely, guided by a learned preference from the non-traditional image subset analysis (Table 8).

• [System Improvement]: Incorporate evaluation metrics that go beyond simple CLIP similarity (RQ2). Develop a Contextual Validity Score that specifically penalizes high alignment scores for images labeled as both traditional and impractical in athletics, shifting the reward function away from simply matching the text to capturing contextual appropriateness.

• [Improved AI System Capability]: The system will move from merely generating images based on text to actively self-correcting its outputs by identifying and neutralizing biased cues introduced during the generation or refinement stages. It will ensure that generated imagery for common activities (e.g., soccer, jogging) adheres strictly to functional appropriateness, regardless of underlying biases in the training data or prompt revision mechanisms.

) 2. Develop a Context-Aware Evaluation Framework for Vision-Language Models (VLMs):

The study shows that standard alignment scores (CLIP, ALIGN) can be misleading because they prioritize nationality cues over activity relevance, leading to systematic misjudgments of stereotypical depictions.

• [System Improvement]: Replace reliance on single alignment metrics with a weighted scoring system that dynamically adjusts the importance of different text elements based on the activity context. For example, when the prompt specifies playing soccer, the weight given to terms related to athletic attire should be prioritized over terms related to cultural identity (country/nationality).

• [System Improvement]: Implement a comparative evaluation mechanism using multiple models (DALL-E 3 vs. Gemini 3 Pro) alongside different scoring methods. If one model consistently assigns higher alignment scores to stereotypical representations, the system should flag that model's output for human review or apply a learned bias correction factor specific to that model's known tendencies (e.g., GPT-4.1 mini's tendency in RQ2).

• [System Improvement]: Create an explicit Construct Validity Check in the evaluation phase, using the findings from RQ2 and Table 6/7. This check will compare alignment scores for traditional vs. non-traditional groups across all activities, ensuring that high alignment is not solely driven by the presence of nationality cues when those cues are contextually irrelevant.

• [Improved AI System Capability]: The system will provide a more trustworthy evaluation layer that judges an image based on its functional relevance to the specified activity, rather than being unduly influenced by learned correlations between nationality and cultural attire embedded in the evaluation models themselves. This leads to a more robust assessment of representational fairness across diverse user prompts.

) 3. Create Dynamically Contextualized Prompt Engineering Strategies:

The analysis of revised prompts (RQ3) revealed that the inclusion of terms like traditional is heavily correlated with higher traditional labels, suggesting prompt rewriting can inadvertently reinforce stereotypes.

• [System Improvement]: Implement a Prompt Revision Strategy that uses semantic clustering (as done in RQ3) to identify and suppress high-frequency, culturally essentializing vocabulary (traditional, specific regional descriptors) when the activity context does not necessitate them.

• [System Improvement]: Develop De-Stereotyping Prompts as an automated countermeasure. If the system detects a high probability of generating an impractical or stereotypical image (e.g., based on prior results), it should automatically generate a prompt variation that explicitly requests contextually appropriate attire, such as: Generate a realistic photograph-like image of a group from [country] playing soccer in a public park, ensuring all individuals are wearing appropriate athletic wear.

• [System Improvement]: Use the thematic clustering (Table 9) to create activity-specific prompt templates. For instance, the Soccer cluster vocabulary should automatically trigger constraints on attire when generating images for that activity, overriding potentially biased general language learned from other activities.

• [Improved AI System Capability]: The system will evolve beyond simple text-to-image generation by employing intelligent prompt engineering that proactively steers the model away from known representational pitfalls, ensuring that user intent (e.g., play soccer) is translated into a depiction of contextually appropriate action, rather than defaulting to culturally exotic tropes.