Cohere researchers conducted a survey of 81 participants across 22 countries to assess the impact of cultural gaps in large language models (LLMs). The study, published on June 23, 2026, found that 83% of participants whose primary language is not English regularly switch to English when using AI, primarily due to higher response quality and accuracy in that language. Data indicates that 89.5% of non-native English speakers feel compelled to switch languages to achieve functional results.

Participants reported significant limitations in AI capabilities regarding cultural awareness. Over one-third of respondents (38%) rated AI’s understanding of their culture below 5 out of 10. The survey highlighted that 63% of users have experienced violations of their cultural norms, such as models defaulting to Western-centric historical narratives, incorrect formality levels, or inappropriate gendered language. These issues are particularly prevalent when users prompt models in languages like Hindi, Punjabi, or regional dialects of French, where systems often fail to capture local context or social nuances.

The consequences of these gaps extend to productivity and marginalization. 67% of participants expressed concern that AI will further marginalize their cultures by reducing them to stereotypes or dominant Western perspectives. This creates a self-reinforcing cycle where users who receive inadequate responses reduce their engagement, limiting the diversity of data available for future model improvement. Users identified creative writing, translation, legal guidance, and local information look-ups as areas requiring the most urgent improvements in cultural sensitivity.

The authors argue that cultural awareness must be a core design requirement in AI systems, moving beyond basic multilingual capabilities. Suggested improvements include region-specific localization that avoids defaulting to dominant cultures, natural handling of code-switched languages (mixing two or more languages in conversation), and the use of diverse sources for historical and social contexts. The researchers emphasize that failing to bridge this gap may lead to a disproportionate distribution of AI-driven productivity gains, benefiting primarily English-speaking, Western-context users while leaving others behind.

Cohere researchers conducted a survey of 81 participants across 22 countries to assess the impact of cultural gaps in large language models (LLMs). The study, published on June 23, 2026, found that 83% of participants whose primary language is not English regularly switch to English when using AI, primarily due to higher response quality and accuracy in that language. Data indicates that 89.5% of non-native English speakers feel compelled to switch languages to achieve functional results.

Participants reported significant limitations in AI capabilities regarding cultural awareness. Over one-third of respondents (38%) rated AI’s understanding of their culture below 5 out of 10. The survey highlighted that 63% of users have experienced violations of their cultural norms, such as models defaulting to Western-centric historical narratives, incorrect formality levels, or inappropriate gendered language. These issues are particularly prevalent when users prompt models in languages like Hindi, Punjabi, or regional dialects of French, where systems often fail to capture local context or social nuances.

The consequences of these gaps extend to productivity and marginalization. 67% of participants expressed concern that AI will further marginalize their cultures by reducing them to stereotypes or dominant Western perspectives. This creates a self-reinforcing cycle where users who receive inadequate responses reduce their engagement, limiting the diversity of data available for future model improvement. Users identified creative writing, translation, legal guidance, and local information look-ups as areas requiring the most urgent improvements in cultural sensitivity.

The authors argue that cultural awareness must be a core design requirement in AI systems, moving beyond basic multilingual capabilities. Suggested improvements include region-specific localization that avoids defaulting to dominant cultures, natural handling of code-switched languages (mixing two or more languages in conversation), and the use of diverse sources for historical and social contexts. The researchers emphasize that failing to bridge this gap may lead to a disproportionate distribution of AI-driven productivity gains, benefiting primarily English-speaking, Western-context users while leaving others behind.