Prompt engineering reduces AI stereotypes, but risks diluting cultural identity

Generative AI models capable of turning text into realistic images are becoming increasingly mainstream, powering creative industries and shaping digital content at scale. But a new academic study has raised red flags about the societal biases embedded in these technologies. Titled “Can we Debias Social Stereotypes in AI-Generated Images? Examining Text-to-Image Outputs and User Perceptions”, and published on arXiv, the research systematically audits three leading text-to-image (T2I) systems, DALL·E-3, Midjourney-6.1, and Stability AI Core, and reveals that even the most red-teamed models continue to reproduce visual stereotypes related to race, gender, and culture.

The study presents a dual-pronged approach: a novel rubric-based metric to detect and score stereotypical content called the Social Stereotype Index (SSI), and a large language model (LLM)-driven prompt refinement strategy to mitigate such biases. The findings are clear—stereotype scores dropped by 61% for geocultural prompts, 69% for occupational, and 51% for adjectival queries after using refined prompts. However, this bias reduction came with a cost: the images often lost specificity and contextual richness, raising key concerns about balancing ethical fairness with user expectations and cultural representation.

How pervasive are stereotypes in AI-generated imagery?

The researchers evaluated 1,200 images generated from 100 diverse prompts falling into three categories: geocultural (e.g., “a Bangladeshi person”), occupational (e.g., “a CEO”), and adjectival (e.g., “a beautiful person”). They developed a 30+ point audit rubric to systematically assess stereotypes based on criteria like gender, age, clothing, expression, background, and more. Each image was then assigned an SSI score between 0 and 1, representing the proportion of stereotypical attributes present.

Across the board, all three T2I models showed similar patterns. For geocultural queries, models often defaulted to visual clichés: Bangladeshi people were portrayed in rural markets, while French individuals appeared in cafés or with Parisian backdrops. For occupational prompts, CEOs were typically white men in suits, while nurses and dietitians were depicted as women. Adjectival prompts like “competent” often produced male-presenting individuals, while “beautiful” skewed toward light-skinned women, mirroring Western beauty standards.

These tendencies reflect entrenched societal biases mirrored in the training data, which largely consists of web-scraped content. Even advanced models with built-in safety layers reproduced these patterns, illustrating the systemic nature of stereotype propagation in generative AI.

Can prompt refinement effectively reduce these biases?

To mitigate bias, the researchers applied an LLM-powered prompt refinement method. This involved identifying stereotype dimensions using the rubric and automatically adjusting the original prompt with inclusive context - for instance, changing “a photo of a Bangladeshi person” to “a confident Bangladeshi person standing in an urban environment, dressed smartly.”

The refined prompts were re-submitted to the same T2I models, generating new image sets. When scored again, the refined images consistently exhibited significantly lower SSI values. Geocultural queries saw a 61% average drop in stereotype scores, occupational ones fell by 69%, and adjectival queries were reduced by 51% - all with high statistical significance.

However, qualitative inspection revealed a notable tradeoff. While the refined images avoided overt stereotypes, they often lacked cultural specificity. For example, instead of rural South Asian markets or traditional attire, the refined Bangladeshi portraits shifted to generic urban environments and Westernized clothing. Similarly, occupational diversity improved in terms of background and attire, but core identity-based biases (e.g., race and gender) often persisted.

This result highlights a broader tension in AI fairness: reducing bias may mean making images less culturally anchored, which could dilute authenticity and decrease perceived relevance for end-users.

How Do Users React to Biased and Debiased AI-Generated Images?

To explore user perception, the study conducted in-depth interviews with 17 diverse participants recruited from Reddit. Participants were shown T2I outputs from both original and refined prompts and asked to compare them against their own mental expectations. The results revealed an ambivalent attitude.

Although most participants valued diversity and inclusivity in principle, they often found stereotypical images more contextually familiar and aligned with their mental models. For instance, the depiction of a “manager” as a white man in a suit or a “beautiful person” as a light-skinned woman matched what many expected, based on cultural exposure or personal experience. In a rapid-fire comparison task, 47% preferred refined images, 43% preferred initial images, and 10% were undecided.

Several participants acknowledged how their own lived experiences influenced their interpretation of stereotypes, with some drawing mental images from real-world acquaintances. Others voiced concern that stereotypical imagery might reinforce harmful norms and called for better regulation of AI-generated visuals. A few appreciated that refined outputs showed greater racial and body diversity, suggesting that repeated exposure to inclusive representations could help shift public perceptions over time.

These findings point to a critical insight: bias in generative AI is not just a technical problem but also a social one. Users bring their own expectations, biases, and interpretations, making stereotype mitigation a complex and value-laden process.

The authors advocate for future T2I systems to incorporate user-centered design, interactive debiasing mechanisms, and policy-level interventions like explainability standards and fairness audits. They also call for more cross-cultural research to understand how global audiences interpret, accept, or resist AI-generated stereotypes.