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

This paper analyzes eight real-world AI deployments across seven non-Western countries to identify six key contextual factors and three influencing dimensions, ultimately proposing 12 guidelines for designing culturally aligned, equitable, and safe AI systems through deep collaboration between developers and domain experts.

The Gist

Imagine you are trying to build a universal translator that helps farmers in rural India, doctors in Kenya, and teachers in Colombia. You might think, "Great! I'll just download the smartest AI model from the internet, plug it in, and it will solve everyone's problems."

This paper says: "Not so fast."

If you try to use a standard, Western-built AI in these non-Western, high-stakes situations, it's like trying to navigate a bustling Mumbai street using a map of rural Vermont. The AI might know the words, but it doesn't know the culture, the rules, or the reality of the people using it.

Here is the story of what the researchers found, explained simply.

The Big Problem: The "One-Size-Fits-All" Trap

The researchers looked at 8 real-world projects (like an AI for farmers, an AI for legal courts, and an AI for health workers) across 7 countries. They found that when you drop a generic AI into a complex, local environment, it often fails. It might give medical advice that violates local laws, translate a legal term incorrectly, or suggest farming techniques that don't work in that specific soil.

The paper argues that to make AI work for "Social Good" (helping people), you can't just be a coder. You have to be a cultural architect.

The 6 Ingredients for Success (The "LISTED" Factors)

The researchers found that successful projects had to juggle six specific ingredients. Think of these as the six legs of a sturdy stool. If one is missing, the whole thing falls over.

1. Language (The Voice): It's not just about translating words. It's about dialects, slang, and how people actually speak.
   • Analogy: Imagine an AI that speaks perfect "Textbook English" but the farmer speaks a local dialect with a heavy accent. The AI hears "plant" but the farmer said "pest." The AI thinks the farmer wants to plant a tree; the farmer is asking how to kill a bug. Successful projects had to teach the AI the local "accent" and even create special dictionaries for local terms.
2. Institution (The Rules): AI has to play by the local rules.
   • Analogy: If you build a car, but the local government only allows tractors on the road, your car is useless. In these projects, the AI had to follow government formats for lesson plans, legal document styles, or health protocols. If the AI didn't match the official paperwork, no one would use it.
3. Safety (The Seatbelt): In high-stakes fields (health, law), a mistake can hurt someone.
   • Analogy: You wouldn't let a self-driving car drive a school bus without a human driver watching. These projects used a "Human-in-the-Loop" approach. The AI does the heavy lifting, but a human expert (like a doctor or judge) double-checks the answer before it goes to the user.
4. Task (The Job): Different jobs need different tools.
   • Analogy: If you are a chef, you need a sharp knife. If you are a painter, you need a brush. An AI designed to summarize news is terrible at diagnosing a disease. These teams had to tweak the AI specifically for the job—sometimes making it slower but more accurate, or changing how it speaks to sound friendly rather than robotic.
5. End-User Demography (The People): Who is using this?
   • Analogy: An app designed for a tech-savvy teenager in New York won't work for an elderly farmer in a village with no internet. Some users can't read, so the AI must speak. Some users are poor, so the AI must work on old phones. The design had to fit the user, not the other way around.
6. Domain (The Expertise): You need a specialist.
   • Analogy: You wouldn't ask a general mechanic to perform brain surgery. Similarly, an AI trained on general internet data doesn't know the specific laws of a local court or the specific diseases of local crops. They needed to feed the AI "curated knowledge" from local experts.

The Three Forces Shaping the Design

Why was this so hard? The paper says three big forces were pulling the strings:

• Sociocultural (The Culture): The AI had to respect local traditions, trust levels, and social norms. If the community didn't trust the AI, they wouldn't use it, no matter how smart it was.
• Institutional (The Organization): Did the local government or NGO have the money and staff to keep the AI running? If the organization was broke or understaffed, the AI project would die, even if the technology was perfect.
• Technological (The Tools): The tools available were often limited. Sometimes the best AI model didn't support the local language at all. Teams had to get creative, using "workarounds" like translating to a dominant language first, then back to the local one.

The Big Surprise: Humans Are the Secret Sauce

The most important finding of the paper is this: Technology alone is not the hero.

In the West, we often think AI will replace human labor. In these non-Western, high-stakes contexts, human labor is what makes the AI work.

• Humans collected the data.
• Humans fixed the translation errors.
• Humans checked the medical advice.
• Humans built the trust with the community.

The researchers found that the most successful projects weren't the ones with the most expensive AI models; they were the ones with the best collaboration between the AI engineers and the local experts (doctors, teachers, farmers).

The 12 Golden Rules (The Takeaway)

The paper ends with 12 guidelines for anyone trying to build AI for good in these settings. Here are the main ones in plain English:

1. Partner Up: Don't build in a silo. Engineers and local experts must work together from day one.
2. Respect the Language: Don't just translate; understand the culture and dialects.
3. Use What You Have: If a local language isn't supported by AI, use a related dominant language but adapt it carefully.
4. Build Trust: People need to trust the system. Sometimes that means showing them a human doctor is checking the AI's work.
5. Follow the Rules: Make sure the AI fits into existing government and organizational workflows.
6. Don't Overpromise: AI can help, but it can't fix broken infrastructure or missing staff.
7. Plan for the Long Haul: Building the AI is just the start. You need money and people to keep it updated forever.
8. Start Small: Test the AI with a small group first to catch mistakes before rolling it out to millions.
9. Be Flexible: Technology changes fast. Build systems that can swap out old models for new ones easily.
10. Test in the Real World: Benchmarks (tests) done in labs are often wrong. Test the AI with real people in real situations.
11. Safety First: Use a mix of simple AI, smart AI, and human checkers to keep people safe.

The Bottom Line

Building AI for social good in non-Western contexts isn't about writing better code. It's about listening better. It's about realizing that a computer program is just a tool, and the real magic happens when that tool is carefully crafted by humans who understand the people it is meant to serve.

The AI is the engine, but the local community and their experts are the steering wheel. Without the steering wheel, the car goes nowhere good.

Technical Summary

Here is a detailed technical summary of the paper "Designing Culturally Aligned AI Systems For Social Good in Non-Western Contexts."

1. Problem Statement

While AI technologies are rapidly expanding into high-stakes domains (education, healthcare, agriculture, law) in non-Western contexts, mainstream AI models often fail to address local realities. These systems frequently suffer from:

• Cultural and Linguistic Misalignment: Outputs are less accurate or relevant for non-Western users, often reproducing Western-centric biases and epistemologies.
• High-Stakes Risks: In domains like healthcare and law, model errors can directly harm livelihoods, health, and legal rights.
• Implementation Gaps: There is a lack of empirical understanding regarding how developers and domain experts actually adapt, design, and deploy these systems in resource-constrained, non-Western environments. The specific trade-offs, required human labor, and sociotechnical infrastructures needed for success remain under-documented.

2. Methodology

The authors conducted a qualitative empirical study to understand the design and deployment of AI for Social Good (AI4SG) in non-Western settings.

• Scope: The study analyzed 8 real-world AI projects spanning 7 countries (India, Bangladesh, Colombia, Kenya, Ethiopia, Nigeria, Ghana) and 18 languages.
• Domains: Education, Healthcare, Agriculture, and Law.
• Data Collection:
  • Semi-structured Interviews: 17 interviews with key informants, comprising 10 AI Developers and 7 Domain Experts.
  • Secondary Research: Analysis of project documentation, model cards, and existing reports.
  • Analysis: Iterative thematic coding using a hybrid approach (inductive themes from data + deductive categories based on AI-adaptation taxonomies).
  • Validation: Member checking was performed with participants to ensure accuracy and interpretive validity.
• Technologies Studied: Machine Translation (MT), Large Language Models (LLMs), Automatic Speech Recognition (ASR), and Text-to-Speech (TTS).

3. Key Contributions

The paper makes three primary contributions:

1. Empirical Account: A grounded analysis of how AI systems are adapted and implemented in high-stakes, non-Western domains.
2. The "LISTED" Framework: A conceptual framework identifying six cross-cutting factors that structure system design, influenced by three higher-order forces.
3. Design Guidelines: 12 actionable guidelines for practitioners to build culturally aligned, equitable, and safe AI systems.

4. Key Results and Findings

A. The "LISTED" Factors

The study identifies six critical factors that shape the design and deployment of culturally aligned AI:

1. Language: Requires grounding models in local linguistic realities. This involves technical adaptations (fine-tuning, prompt engineering) and community-built data. Challenges include low-resource languages, dialects, and code-mixing.
2. Institution: Adoption depends on alignment with government regulations, organizational workflows, and institutional priorities (e.g., lesson plan formats, legal protocols). Institutional trust is often a prerequisite for scale.
3. Safety: Relies heavily on Human-in-the-Loop (HIL) workflows. Technical safeguards (guardrails, classifiers) are insufficient alone; human oversight is critical for high-stakes decisions, especially in multilingual contexts where safety guardrails may fail.
4. Task: Specific task constraints (latency, modality, environmental noise) dictate technical choices. For example, ASR for reading assessments must preserve errors rather than correct them, and rural field conditions require noise-robust models.
5. End-User Demography: Literacy levels, age, gender, and income dictate interaction modalities (voice vs. text) and access channels (SMS vs. WhatsApp). Systems must adapt to local cultural norms and symbols.
6. Domain: Reliability is anchored by domain-specific knowledge. Curated knowledge bases (RAG) are often preferred over general LLM world knowledge to ensure accuracy in specialized fields like law and medicine.

B. Higher-Order Influences

These six factors are shaped by three overarching influences:

• Sociocultural: Diversity of practices, linguistic nuances, and trust dynamics.
• Institutional: Resources, policies, and the capacity of local organizations to sustain systems.
• Technological: Capabilities and limits of available models (e.g., lack of support for specific African languages).

C. The Critical Role of Human Labor

A central finding is that human labor is more critical than technological expertise alone for achieving safe and effective outcomes.

• Collaboration: Successful systems required deep, ongoing collaboration between AI developers and domain experts.
• Adaptation Effort: Significant human effort was required for data curation, annotation, creating glossaries, and continuous model oversight.
• Sustainability: Long-term viability depends on institutional capacity to fund and manage this human labor, not just the initial technical build.

D. Adaptation Strategies

The study details the spectrum of technical adaptations used:

• In-Context Adaptations: Prompt engineering, retrieval-augmented generation (RAG), and glossaries. These are lower-cost and faster to implement.
• In-Weight Adaptations: Fine-tuning, continual pre-training, and training new models. These are resource-intensive but necessary for low-resource languages or specific domain requirements.
• Hybrid Approaches: Many projects used a mix, such as using English for generation and local models for translation, or using classifiers to route queries to humans.

5. Significance and Guidelines

The paper argues that building AI for social good in non-Western contexts requires moving beyond "techno-solutionism." It emphasizes that cultural alignment is a sociotechnical process, not just a technical one.

The authors provide 12 Design Guidelines categorized by influence:

• Sociocultural (G1-G4): Emphasize equal partnership between developers and local experts; address linguistic gaps; utilize dominant languages as bridges for unsupported dialects; and design for community trust (mitigating both skepticism and over-reliance).
• Institutional (G5-G8): Align with institutional mandates and workflows; recognize AI cannot substitute for missing infrastructure; ensure sustainable institutional capacity for long-term maintenance; and prioritize in-context methods before heavy in-weight investments based on capacity.
• Technological (G9-G12): Design flexible systems to accommodate rapid model shifts; use small-scale deployments to expose real-world risks; develop localized evaluation metrics beyond standard benchmarks; and implement layered safety (lightweight models + HIL + frontier models).

Conclusion:
The paper concludes that effective AI systems in non-Western, high-stakes contexts are not built solely on superior algorithms. They are the result of extensive human labor, deep institutional alignment, and iterative cultural adaptation. The "LISTED" framework and the 12 guidelines offer a roadmap for practitioners to navigate the complex trade-offs between technical feasibility, cultural relevance, and ethical safety.