Designing Culturally Aligned AI Systems For Social Good in Non-Western Contexts

Paper Title

Designing Culturally Aligned AI Systems For Social Good in Non-Western Contexts

Publication Info

• Topic area: AI system design for social good in non-Western, high-stakes domains.
• Keywords: AI for social good, cultural alignment, non-Western contexts, high-stakes domains, language adaptation, human-in-the-loop, institutional capacity, safety mechanisms, sociotechnical systems.

Background and Problem

• Problem / challenge: Existing AI systems often fail to address the sociocultural, linguistic, and institutional complexities of non-Western contexts, leading to biases, lack of cultural awareness, and reduced effectiveness in high-stakes domains.
• Significance: Addressing these gaps is critical for ensuring equitable, safe, and effective AI deployments that can positively impact education, healthcare, agriculture, and law in underrepresented regions.
• Motivation and related work: Prior research highlights the risks of AI systems reproducing biases, lacking cultural alignment, and entrenching colonial epistemologies. While calls for culturally responsive AI have been made, there is limited empirical understanding of how to design and implement such systems in practice. This study builds on these gaps by analyzing real-world deployments.

Solution

• Proposed approach: The study identifies six cross-cutting factors (Language, Institution, Safety, Task, End-User Demography, Domain — LISTED) and three higher-level influences (Sociocultural, Institutional, Technological) that shape the design and deployment of culturally aligned AI systems.
• Novelty:
  1. Empirical analysis of eight real-world AI deployments across seven countries and 18 languages.
  2. A framework detailing six factors and three overarching influences for AI system design in non-Western contexts.
  3. Twelve actionable guidelines for practitioners to build culturally aligned, socially beneficial AI systems.
• Procedure and key techniques:
  • Semi-structured interviews with 17 AI developers and domain experts.
  • Secondary research on project documentation and reports.
  • Thematic analysis to identify patterns and challenges in AI adaptation.
  • Examination of adaptation strategies, including in-context and in-weight approaches.

Results

• Concrete findings:
  • Six factors (LISTED) consistently shaped AI system design: Language adaptation, institutional alignment, safety mechanisms, task-specific constraints, end-user demographic considerations, and domain-specific knowledge.
  • Human labor was central to ensuring cultural alignment, safety, and contextual relevance.
  • Adaptation strategies ranged from lightweight (e.g., prompt engineering) to resource-intensive (e.g., fine-tuning, data collection).
• Advantage over baselines:
  • Systems grounded in local languages, institutional rules, and domain-specific knowledge outperformed generic AI models in usability, trust, and adoption.
  • Human-in-the-loop workflows mitigated risks and ensured reliability in high-stakes scenarios.
• Experiments / evaluation:
  • Analysis of eight projects across education, healthcare, agriculture, and law.
  • Deployment scales ranged from pilot (<500 users) to large-scale (~350,000 users).
  • Evaluation metrics included user feedback, task-specific performance, and safety outcomes.
• Limitations and future work:
  • Limited focus on long-term sustainability and scalability of human-in-the-loop systems.
  • Need for further exploration of hyper-local cultural variations and their integration into AI systems.
  • Future work should address infrastructural disparities and expand the guidelines to other domains and contexts.

Summary

This paper explores how AI systems can be designed for social good in non-Western, high-stakes contexts by analyzing eight real-world deployments. It identifies six key factors (LISTED) and three overarching influences (Sociocultural, Institutional, Technological) that shape system design and deployment. The study highlights the critical role of human labor, collaboration between AI developers and domain experts, and the use of both lightweight and resource-intensive adaptation strategies. The findings are synthesized into 12 actionable guidelines to support practitioners in building culturally aligned, equitable, and responsive AI systems. These insights are applicable to diverse domains, offering a roadmap for creating effective AI solutions in underrepresented regions.