``It Hasn’t Lived in Our Society”: Investigating Cultural Sensitivity in LLM Chatbots for Emotional Support

Paper Title

"It Hasn’t Lived in Our Society”: Investigating Cultural Sensitivity in LLM Chatbots for Emotional Support

Publication Info

• Topic area: Cultural adaptation of Large Language Models (LLMs) for emotional support in underrepresented contexts.
• Keywords: Large Language Models, cultural sensitivity, emotional support, Saudi Arabia, mental well-being, chatbots, human-computer interaction, multicultural design, value-sensitive design, mental health.

Background and Problem

• Problem / challenge: Generic LLMs, such as ChatGPT, lack cultural sensitivity, particularly in non-Western contexts, leading to advice that may conflict with local norms, values, and beliefs. This limits their effectiveness and may cause harm in culturally specific emotional support scenarios.
• Significance: Addressing cultural sensitivity in LLMs can enhance their relevance, safety, and acceptance, especially in underrepresented populations like young Saudi women, who face unique sociocultural challenges and mental health stigma.
• Motivation and related work: Prior studies have shown the potential of LLMs for emotional support but highlighted their limitations in non-WEIRD (Western, Educated, Industrialized, Rich, Democratic) contexts. Existing research has not adequately addressed cultural alignment in emotional support tools, particularly for Arab populations. This paper builds on frameworks like the Ecological Validity Model (EVM) and multicultural guidelines to address this gap.

Solution

• Proposed approach: Development of the Culturally Sensitive Emotional Support Chatbot (CSESC), a technology probe tailored to the Saudi cultural context, integrating cultural, religious, and social norms into emotional support responses.
• Novelty:
  1. Examining the cultural sensitivity of generic LLMs and emotional support frameworks.
  2. Designing culturally sensitive, empirically grounded prompts for emotional support.
  3. Identifying limitations of culturally adapted LLMs in emotional support tasks.
  4. Introducing the concept of "minimum cultural alignment" (MCA) for culturally sensitive design.
• Procedure and key techniques:
  1. Selection of Arabic-capable baseline models (e.g., GPT-4, Gemini, ALLaM).
  2. Assessment of cultural alignment using three prompt configurations: generic GPT, GPT with Emotional Support Conversation (ESC) prompts, and GPT with ESC plus Hofstede’s cultural dimensions.
  3. Development of CSESC prompts based on the EVM and local expert input.
  4. Evaluation of CSESC through expert comparison with therapist-authored responses and user studies with 21 young Saudi women.

Results

• Concrete findings:
  • CSESC responses aligned with therapist-authored responses in 78% of cases, compared to 3% for generic GPT.
  • CSESC was preferred by users in 87% of cases over generic GPT.
  • Participants valued cultural alignment, including family dynamics, religious references, and sensitivity to Saudi social norms.
• Advantage over baselines:
  • Generic GPT lacked cultural, social, and religious dimensions, often providing advice that conflicted with Saudi norms.
  • CSESC demonstrated higher cultural appropriateness, religious sensitivity, and alignment with social norms.
• Experiments / evaluation:
  • Expert evaluation compared CSESC and GPT responses against therapist-authored benchmarks.
  • User study with 21 young Saudi women assessed perceptions of cultural alignment, emotional support, and usability.
  • Methods included semi-structured interviews, role-playing with personas, and side-by-side response comparisons.
• Limitations and future work:
  • Findings are specific to the Saudi context and young, educated women.
  • Short-term evaluation limits insights into long-term use.
  • Future work should explore diverse populations, real-world deployment, and evolving LLM capabilities.

Summary

This paper addresses the lack of cultural sensitivity in LLM-based emotional support tools by developing and evaluating the Culturally Sensitive Emotional Support Chatbot (CSESC) tailored to the Saudi context. CSESC demonstrated significant improvements in cultural alignment, religious sensitivity, and social norm awareness compared to generic LLMs like ChatGPT. User studies highlighted the importance of culturally grounded design for enhancing trust, relevance, and emotional support. The study introduces the concept of "minimum cultural alignment" (MCA) as a baseline for culturally sensitive design, offering a framework for adapting LLMs to other cultural settings.