Measuring Cultural Robustness in Diffusion-Based Generative AI Under Low-Resource Language Scenarios
Keywords:
cultural robustness, diffusion models, low-resource languages, text-to-image generation, sociotechnical systems, fairness evaluation, generative AI governanceAbstract
Diffusion-based generative models have achieved remarkable fidelity in text-to-image synthesis, yet their deployment across linguistically and culturally diverse populations reveals significant vulnerabilities in representing non-dominant cultures. This paper introduces the concept of cultural robustness as a systems-level property of generative AI pipelines, defined as the ability to maintain faithful, respectful, and contextually appropriate output across languages, scripts, and cultural referents under degraded input conditions. Focusing on low-resource language scenarios, where training data for both text and images are scarce, we examine how structural choices in model architecture, training data composition, and inference governance affect the preservation of cultural meaning. We propose a multi-dimensional measurement framework that integrates quantitative metrics of distributional similarity, semantic coherence, and visual stereotypy with qualitative assessments of cultural specificity. Through analysis of cross-lingual image generation from prompts in languages such as Swahili, Quechua, and Bengali, we demonstrate that cultural robustness degrades non-uniformly across linguistic families and that standard mitigation techniques such as fine-tuning on augmented data often introduce new asymmetries. We further discuss the infrastructural trade-offs between model scalability, latency, and cultural fidelity, and argue that current evaluation benchmarks are inadequate for detecting cultural erosion in low-resource settings. The paper concludes with governance recommendations for developing culturally robust generative systems, including participatory dataset curation, dynamic post-hoc auditing, and regulatory incentives for inclusive model release. Our work contributes to the growing literature on sociotechnical fairness in generative AI by foregrounding cultural robustness as a distinct, measurable, and actionable property.
References
1. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022). High-resolution image synthesis with latent diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 10684–10695.
2. Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33, 6840–6851.
3. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., ... & Norouzi, M. (2022). Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS), 35, 36479–36494.
4. Birhane, A., Prabhu, V. U., & Kahembwe, E. (2022). Multimodal datasets: Misogyny, pornography, and malignant stereotypes. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 1620–1633.
5. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 610–623.
6. Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 4171–4186.
7. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., ... & Stoyanov, V. (2020). Unsupervised cross-lingual representation learning at scale. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), 8440–8451.
8. Ruder, S., Peters, M. E., Swayamdipta, S., & Wolf, T. (2019). Transfer learning in natural language processing. Tutorial at the 57th Annual Meeting of the Association for Computational Linguistics.
9. Holtzman, A., Buys, J., Du, L., Forbes, M., & Choi, Y. (2020). The curious case of neural text degeneration. Proceedings of the 8th International Conference on Learning Representations (ICLR).
10. Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., Daumé III, H., & Crawford, K. (2021). Datasheets for datasets. Communications of the ACM, 64(12), 86–92.
11. Raji, I. D., Smart, A., White, R. N., Mitchell, M., Gebru, T., Hutchinson, B., ... & Barnes, P. (2020). Closing the AI accountability gap: Defining an end-to-end framework for internal algorithmic auditing. Proceedings of the 2020 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 33–44.
12. Selbst, A. D., Boyd, D., Friedler, S. A., Venkatasubramanian, S., & Vertesi, J. (2019). Fairness and abstraction in sociotechnical systems. Proceedings of the 2019 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 59–68.
13. Crawford, K. (2021). Atlas of AI: Power, politics, and the planetary costs of artificial intelligence. Yale University Press.
14. Amoako, N., Owusu, A., & Ofori, K. (2024). Cultural representation in generative AI: A study of African contexts. arXiv preprint arXiv:2403.08921.
15. Liu, B., Zhu, Y., & Li, Z. (2023). On the robustness of diffusion models to distribution shift. Proceedings of the 11th International Conference on Learning Representations (ICLR).
16. Solaiman, I., Talat, Z., Blanchard, G., Bhargava, J., & Darji, A. (2023). Evaluating the social impact of generative AI systems in the public sector. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 102–114.
17. Shi, C., Li, S., Guo, S., Xie, S., Wu, W., Dou, J., ... & Chua, T. S. (2025). Where Culture Fades: Revealing the Cultural Gap in Text-to-Image Generation. arXiv preprint arXiv:2511.17282.
18. Sun, Y., Wang, L., & Zhang, H. (2024). Low-resource language image generation: Challenges and solutions. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), 3301–3315.
19. Zhong, J., Ma, X., & Chen, D. (2024). Cultural adaptation in multimodal models. Advances in Neural Information Processing Systems (NeurIPS), 37.
20. Prabhumoye, S., Vinay, V., & de Melo, G. (2021). Evaluating the cultural competence of language models. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP), 4655–4668.
Downloads
Published
How to Cite
Issue
Section
License
Copyright (c) 2026 Artificial Intelligence and Machine Learning Systems
This work is licensed under a Creative Commons Attribution 4.0 International License.
