Leading AI models, including Claude Sonnet 4.5, GPT-5.4, and Gemini 2.5 Flash, consistently exhibit a Western-individualist bias in their advice, even when interacting with users from collectivist cultures. This systemic homogenization of values suggests AI systems may not adequately reflect diverse global perspectives, potentially reinforcing existing cultural norms rather than adapting to local values.
| Attribute | Value |
|---|---|
| Released by | arXiv cs.CL |
| Release date | |
| What it is | A cross-cultural audit of individualism-collectivism bias in leading large language models. |
| Who it is for | AI developers, ethicists, researchers, and policymakers. |
| Where to get it | arXiv (2604.22153) |
| Price | Free |
- AI models show a significant Western-individualist bias in advice across cultures.
- Claude Sonnet 4.5, GPT-5.4, and Gemini 2.5 Flash were tested.
- The bias is measured against World Values Survey data.
- Nigeria and India show the largest gaps in AI advice versus local values.
- AI systems may encode outdated stereotypes, as seen with Japan.
- AI systems consistently provide Western-style, individualist advice, even to users from collectivist societies [1].
- This bias creates a mean gap of +0.76 on a 1-5 scale compared to local values [1].
- The largest bias gaps were observed for Nigeria (+1.85) and India (+0.82) [1].
- AI models like Claude and GPT-5.4 exhibit nearly identical bias magnitudes [1].
- Gemini shows a lower but still significant bias magnitude [1].
- AI may encode outdated stereotypes, treating Japanese users as more group-oriented than surveys indicate [1].
- Generative AI models can reflect and amplify cultural bias present in their training data [4].
What is AI Cultural Bias?
AI cultural bias refers to the phenomenon where artificial intelligence systems reflect and amplify cultural norms or stereotypes present in their training data. Generative AI models can associate certain professions with specific genders if such patterns are prevalent in the data [4]. This can lead to AI providing advice or generating content that aligns with one cultural perspective over others [1]. UNESCO works to ensure AI’s development protects and promotes cultural diversity [7].
What is new vs the previous version?
This research rigorously measures individualism-collectivism bias in leading large language models, providing specific quantitative data on its magnitude and mechanisms. Previous work has proposed frameworks to fix cultural bias, but this study focuses on upstream measurement [2]. It identifies how different AI models respond to signals like language versus explicit identity [2]. The study also reveals that AI can encode outdated stereotypes, as seen in its treatment of Japanese users [1].
How does this cultural bias work in AI?
Cultural bias in AI works by reflecting patterns and values embedded in the vast datasets used for training large language models. AI systems were presented with ten real-life personal dilemmas framed for users from 10 countries in 7 languages [1]. The AI advice was compared against World Values Survey data on what people in each country believe [1]. All three AI systems consistently gave Western-style, individualist advice [1]. Claude shifts further collectivist in the user’s native language [1]. Gemini shifts more individualist, and GPT-5.4 responds only to stated country identity [1].
Benchmarks and evidence
| AI System | Bias Magnitude (1-5 scale) | Mechanism of Bias | Source |
|---|---|---|---|
| Claude Sonnet 4.5 | Nearly identical to GPT-5.4 | Shifts further collectivist in user’s native language | [1] |
| GPT-5.4 | Nearly identical to Claude Sonnet 4.5 | Responds only to stated country identity | [1] |
| Gemini 2.5 Flash | Lower but still significant | Shifts more individualist | [1] |
| Mean Gap (all AI) | +0.76 (t=15.65, p<0.001) | Consistent Western-individualist advice | [1] |
| Largest Gap (Nigeria) | +1.85 | AI advice significantly more individualist than local values | [1] |
| Largest Gap (India) | +0.82 | AI advice significantly more individualist than local values | [1] |
| Japan (exception) | Not disclosed | AI treated users as more group-oriented than surveys show | [1] |
Who should care about AI cultural bias?
Builders
AI developers and engineers should care about cultural bias to build more diverse and inclusive AI teams [6]. Cross-functional teams with engineers, social scientists, and ethicists can spot unintended biases [3]. Model cards listing intended uses, training data, and bias risks could be mandated for AI models [3].
Enterprise
Companies deploying AI should be aware of cultural bias to avoid perpetuating stereotypes and ensure their AI tools are fair. Algorithmic bias becomes objectionable when it systematically disadvantages unprivileged groups [6]. Businesses can minimize algorithmic bias across data collection and model design [6].
End users
End users should be aware that AI advice may not align with their cultural values, especially when seeking guidance on personal matters. AI systems consistently give Western-style, individualist advice [1]. This could lead to advice that conflicts with family, community, or authority priorities in collectivist societies [1].
Investors
Investors should consider the ethical implications and potential market limitations of AI systems exhibiting cultural bias. AI that fails to protect and promote cultural diversity may face regulatory scrutiny or limited global adoption [7]. Investing in AI fairness and bias mitigation can lead to more robust and widely accepted products.
Risks, limits, and myths
- Risk: Homogenization of values: AI systems may systematically homogenize values, promoting a Western-individualist perspective globally [1].
- Risk: Reinforcing stereotypes: AI can encode outdated stereotypes, as seen in its treatment of Japanese users [1].
- Risk: Widening gender gaps: New technologies risk widening gender gaps, especially when AI learns from biased data [7].
- Limit: Training data bias: Generative AI models can reflect and amplify cultural bias present in their training data [4].
- Myth: AI is culturally neutral: This study demonstrates that AI is not culturally neutral; it exhibits significant biases [1].
- Myth: Language alone dictates cultural response: While language plays a role, GPT-5.4 responds only to stated country identity, showing other factors are at play [1].
FAQ
- What is individualism-collectivism in AI research?
- Individualism-collectivism is a common dichotomy in cross-cultural research, used to frame how AI systems reflect cultural values [2, 5]. It captures aspects like family duty, authority deference, and attitudes towards divorce [2].
- Which AI models were tested for cultural bias?
- The study tested Claude Sonnet 4.5, GPT-5.4, and Gemini 2.5 Flash for cultural bias [1]. These are leading AI systems in the field.
- How was AI cultural bias measured?
- AI cultural bias was measured by presenting AI systems with personal dilemmas framed for users from 10 countries in 7 languages [1]. AI advice was then compared against World Values Survey data on actual beliefs in each country [1].
- Which countries showed the most significant AI cultural bias?
- Nigeria showed the largest gap in AI advice versus local values (+1.85), followed by India (+0.82) [1]. This indicates a strong Western-individualist bias in AI responses.
- Does AI always give individualist advice?
- AI systems consistently gave Western-style, individualist advice, but Japan was an exception [1]. AI treated Japanese users as more group-oriented than surveys show, indicating outdated stereotypes [1].
- How do different AI models handle cultural signals?
- Claude shifts further collectivist in the user’s native language, Gemini shifts more individualist, and GPT-5.4 responds only to stated country identity [1]. This shows models diverge in their mechanisms.
- Can cultural bias in AI be fixed?
- Addressing cultural bias requires rigorous measurement, diverse development teams, and careful model design [2, 3, 6]. Strategies include minimizing algorithmic bias and building inclusive teams [6].
- Why is cultural diversity important for AI development?
- Cultural diversity in AI development is crucial to prevent the homogenization of values and ensure AI protects and promotes diverse cultures [1, 7]. Diverse teams can identify and mitigate unintended biases [3].
Glossary
- Individualism-Collectivism
- A cultural dimension describing the degree to which individuals are integrated into groups, with individualism emphasizing personal goals and collectivism emphasizing group goals [5].
- Large Language Model (LLM)
- An artificial intelligence program trained on vast amounts of text data, capable of understanding, generating, and translating human language [1].
- Algorithmic Bias
- Systematic and repeatable errors in a computer system that create unfair outcomes, such as favoring privileged groups or disadvantaging unprivileged ones [6].
- World Values Survey (WVS)
- A global research project that explores people’s values and beliefs, providing data on cultural norms across countries [1].
- Generative AI
- Artificial intelligence that can create new content, such as text, images, or audio, based on patterns learned from its training data [4].
Sources
- [1] [2604.22153] When AI Speaks, Whose Values Does It Express? A Cross-Cultural Audit of Individualism-Collectivism Bias in Large Language Models — https://arxiv.org/abs/2604.22153
- [2] When AI Speaks, Whose Values Does It Express? A Cross-Cultural Audit of Individualism–Collectivism Bias in Large Language Models — https://arxiv.org/html/2604.22153
- [3] AI Bias: Cultural Stereotypes in AI Writing — https://blog.writetheworld.org/ai-bias-cultural-stereotypes
- [4] Generative AI – Wikipedia — https://en.wikipedia.org/wiki/Generative_artificial_intelligence
- [5] Individualism – Wikipedia — https://en.wikipedia.org/wiki/Individualism
- [6] What Is Artificial Intelligence (AI)? | IBM — https://www.ibm.com/think/topics/artificial-intelligence
- [7] Artificial Intelligence and emerging technologies | UNESCO — https://www.unesco.org/en/artificial-intelligence
