Cultural Perspectives and Expectations for Generative AI: A Global Survey Approach

This paper presents findings from a large-scale global survey that explores diverse cultural perspectives on Generative AI, distilling community-defined understandings of culture to propose recommendations for more inclusive and sensitive AI development, including participatory approaches and frameworks for addressing cultural boundaries.

The Gist

Imagine you are a chef trying to cook a single, giant pot of stew for the entire world. You want everyone to enjoy it, but you realize that what tastes like "comfort food" in one country might be considered offensive or even inedible in another.

This paper is essentially a massive taste-test survey conducted by Google researchers to figure out exactly what ingredients (cultural elements) people want in their AI "stew" and, more importantly, which ingredients they absolutely refuse to eat.

Here is the breakdown of their findings, translated into everyday language:

1. The Problem: The AI Chef is Missing the Recipe

Right now, Generative AI (the tech that writes stories or draws pictures) is like a chef who only learned to cook using recipes from a few specific neighborhoods (mostly Western, English-speaking ones). Because of this, when the AI tries to cook a dish for someone in India, Nigeria, or Brazil, it often gets the spices wrong. It might accidentally serve a "sacred" dish as a "joke" or mix up historical facts, leading to stereotypes or hurt feelings.

The researchers asked: "How do we teach this AI to respect the unique flavors of every culture?"

2. The Survey: Asking 5,600 People What Matters

To answer this, they didn't just guess. They asked 5,629 people from 13 different countries (including the US, Brazil, India, Japan, Nigeria, and more) two main things:

• "What does 'culture' mean to you?"
• "What parts of your culture should the AI NEVER touch?"

3. The Big Discoveries (The "Flavor Profiles")

A. What is Culture? (It's not just geography)
People didn't just say, "My culture is my country." They described it as a mix of:

• The "Ancestral" Flavor: For many in Asia and Africa, culture is about ancestors, traditions, and the "total way of life."
• The "Artistic" Flavor: In Europe, people often pointed to specific art, music, and famous landmarks (like the Eiffel Tower).
• The "Religious" Flavor: This was the big winner. Across almost every country, Religion and Tradition were the most important parts of identity.

B. The "Red Lines" (The Forbidden Ingredients)
The researchers asked: "Is there anything you want the AI to never draw or talk about?"

• The Consensus: Over 20% of people said "Yes," and in some countries, it was over 30%.
• The Top Forbidden Items:
  • Sacred Things: Prayers, religious rituals, and holy texts. People feel AI lacks the "soul" to handle these respectfully.
  • Trauma: Specific historical tragedies (like slavery in the US or the Holocaust in Germany) that shouldn't be turned into a quick AI image.
  • Fake People: Many people (especially in Germany) said, "Don't make fake pictures of real people at all."
  • Misused Symbols: Using sacred symbols (like a religious statue) just to sell a product or make a meme.

C. The "Surprise" Sensitivities
While religion was the global "Red Line," different countries had their own specific sensitivities:

• South Korea: People were very sensitive about health status and where you live.
• India & UAE: The concept of "Caste" (a social hierarchy) was a major sensitive topic, even though it wasn't a top concern in other countries.
• The US & Germany: Historical trauma was a huge "do not touch" zone.

4. The Recommendations: How to Fix the AI Kitchen

The authors suggest that AI developers stop trying to use a "one-size-fits-all" recipe. Instead, they propose a new approach with four pillars:

1. Listen First (Awareness): Don't just guess what culture is. Use surveys and interviews to actually ask people what matters to them.
2. Invite the Locals (Participation): Instead of just having a team of engineers in California decide what's "safe," bring in local religious leaders, historians, and community members to help train and check the AI. Think of it as hiring a local taste-tester for every region.
3. Customize the Settings (Multi-facetedness): The AI should have "regional settings."
   • Example: If a user in the UAE asks for an image of a religious site, the AI should be extremely strict and accurate. If a user in France asks for a painting of a historical event, the AI might have more creative freedom.
4. Create "Safety Layers" (Nuance): Not all sensitive topics are the same. The authors suggest a tiered system:
   • Tier 1 (The Hard Stop): Things that should never be generated (e.g., specific prayers, sacred rituals).
   • Tier 2 (The High-Stakes Zone): Things that can be generated, but only if the AI is 100% accurate and doesn't stereotype (e.g., historical figures, religious artifacts).

The Bottom Line

This paper argues that for AI to truly be "global," it needs to stop acting like a tourist who only knows the surface-level facts and start acting like a respectful guest who understands the deep, sacred, and sometimes painful history of the people it is serving.

In short: If you want to build an AI that the whole world trusts, you have to ask the world what they are comfortable with, respect their "Red Lines," and realize that Religion and Tradition are the most important ingredients in the recipe.

Technical Summary

Here is a detailed technical summary of the paper "Cultural Perspectives and Expectations for Generative AI: A Global Survey Approach."

1. Problem Statement

Generative AI (GenAI) systems are increasingly deployed globally but suffer from significant gaps in cultural representation and understanding. Current models often reflect "WEIRD" (Western, Educated, Industrialized, Rich, Democratic) biases, leading to the stereotyping or erasure of non-Western cultures. This is exacerbated by:

• Training Data Limitations: Models are trained on majority-English internet datasets lacking cultural nuance.
• Lack of Expertise: AI researchers often lack training in anthropology or regional cultural studies.
• Absence of Metrics: Major AI firms lack standardized metrics for "cultural competency," leaving benchmarks at a research stage.
• The "Inverse Problem": As GenAI gains the ability to represent culturally sensitive imagery at scale, there is no empirical evidence regarding what specific cultural elements global populations consider sensitive, sacred, or off-limits ("redlines").

The paper aims to address this asymmetry by providing empirical data on how diverse global populations define culture and what boundaries they expect GenAI to respect.

2. Methodology

The study employed a large-scale, cross-sectional survey conducted between September 2024 and January 2025.

• Participants: A total of 5,629 respondents from 13 countries (Brazil, Cameroon, France, Germany, India, Indonesia, Italy, Japan, Mexico, Nigeria, South Korea, UAE, and the USA).
• Demographics: The sample was quota-stratified for gender (approx. 50/50 male/female) and age (18+). The survey was administered in local dominant languages (e.g., Portuguese, French, German, Bahasa, Japanese, Arabic) to minimize barriers.
• Survey Instrument:
  • Developed through a review of cultural values literature and AI ethics, validated by subject matter experts in linguistics, anthropology, and HCI.
  • Structure: Included 24 items, with the analysis focusing on 8 key areas:
    1. Open-ended definitions: "What is your understanding of the word culture?"
    2. Social Identity Prioritization: Selecting the top 3 important identities from a list of 19 (e.g., religion, caste, nationality).
    3. Important Cultural Artifacts: Open-ended descriptions of key artifacts/practices.
    4. GenAI Familiarity: Likert scale assessment of user knowledge.
    5. Cultural Redlines: Binary question on whether any cultural aspects should never be represented by AI, followed by open-text elaboration.
    6. Sensitivity Ranking: Ranking cultural categories (based on UNESCO frameworks) and social identities by perceived sensitivity.
• Data Analysis:
  • Quantitative: Weighted to reflect local population characteristics and global proportions (using World Bank data). Analyzed using Python (Pandas, Numpy, Seaborn).
  • Qualitative: Inductive thematic analysis of open-ended responses, translated and reviewed for accuracy.

3. Key Contributions

• Empirical Baseline: Provides the first large-scale, multi-national dataset quantifying global attitudes toward GenAI cultural representation.
• Operational Definitions of Culture: Moves beyond abstract definitions to distill "working definitions" of culture directly from community responses, highlighting regional variations (e.g., tangible artifacts in Europe vs. collective heritage in Asia).
• Identification of "Cultural Redlines": Establishes specific boundaries where users believe AI representation is unacceptable, moving beyond generic safety filters to culturally specific prohibitions.
• Framework for Sensitivity: Proposes a tiered taxonomy for handling cultural content, distinguishing between "sensitive" (requiring high accuracy) and "prohibitive" (requiring non-generation) content.

4. Key Results

A. Understanding of Culture

• Global Consensus: Culture is generally viewed as a composite of language, traditions, beliefs, and artistic expressions.
• Regional Nuances:
  • Europe: Focused on tangible artifacts (music, literature, history, architecture).
  • Asia: Focused on collective heritage (ancestors, nation, country).
  • Sub-Saharan Africa: Defined culture as a "total way of life" and "collections of attitudes."

B. Social Identities: Importance vs. Sensitivity

• Top Identity: Religion and Tradition was the most consistently cited identity (48% globally) and showed the highest stability between being "important" and "sensitive" (65.6% agreement).
• The "Redline" Effect: Certain identities showed a "low-importance, high-sensitivity" pivot. For example, Caste was ranked low in general importance globally but had a 64.6% sensitivity agreement rate. Military/Veteran Status had the highest sensitivity agreement (74.5%).
• Ascribed vs. Achieved: Ascribed identities (Ethnicity, Caste, Religion) were consistently more sensitive than achieved identities (Occupation, Education).

C. GenAI Familiarity and Perceptions

• Familiarity: 68% of respondents were familiar with GenAI; 59% had actively used it. Familiarity was highest in South Korea and the UAE, lowest in France and Cameroon.
• Primary Threat: Regardless of familiarity level, "inaccurate information" was perceived as the primary threat to cultural representation.
• Ambivalence: Approximately 50% of respondents expressed ambivalence about GenAI's ability to represent culture, citing a lack of intrinsic human emotion and creativity.

D. Cultural Redlines (What should NEVER be represented)

• Prevalence: Over 20% of participants globally (reaching 30%+ in many countries) stated that certain aspects of their culture should never be represented by GenAI.
• Specific Prohibitions:
  • Religion/Sacredness: Religious artifacts, prayers, sacred texts, and rituals were the most common "redlines."
  • Historical Trauma: Specific events (e.g., slavery in the US, the Holocaust in Germany, massacres in Indonesia) were flagged as inappropriate for AI generation.
  • Human Dignity: Concerns about "fake images" of people and the distortion of identity led some to suggest no human depiction at all.
  • Commercialization: In Brazil, commercializing cultural elements like Samba or Carnival without context was cited as a redline.

5. Significance and Recommendations

The paper argues that current "one-size-fits-all" safety filters are insufficient for global GenAI deployment. The authors propose a four-pillar methodological approach for developers:

1. Awareness: Integrate empirical surveys and user perspectives as a standard step in the development cycle, not just an afterthought.
2. Participation: Move beyond passive data collection to active user participation via:
   • Dimension-Based Judge Models: Using raters from specific cultural/religious groups for RLHF.
   • Community-Led Norms: Defining "sacred" and "sensitive" with community custodians.
   • Collaborative Audits: Post-deployment auditing with community engagement.
3. Multi-facetedness: Implement Dynamic Configuration (e.g., region-specific "cultural configuration files") and Contextual Adaptation (adjusting model "temperature" based on local sensitivity to identity vs. art).
4. Nuance (Tiered Sensitivity Framework):
   • Tier 1 (Prohibitive Redline): Sacred rituals, prayers, and divine depictions that should never be generated.
   • Tier 2 (High-Fidelity Representation): Cultural artifacts where generation is allowed only if historical/theological accuracy is guaranteed (e.g., via RAG with authoritative sources).

Conclusion: The study concludes that religion and tradition are the most critical dimensions for GenAI safety. Ignoring these specific, culturally grounded "redlines" risks perpetuating colonial dynamics and causing significant harm. The paper calls for a shift from geographic generalizations to a nuanced, dimension-based approach to cultural alignment in AI.