Large Language Models (LLMs) are rapidly being integrated into educational settings, yet fundamental questions about their cultural biases and perceptual frameworks remain unexplored. Our study examines how different LLMs exhibit distinct “cultural identities” when responding to ambiguous classification tasks.

The investigation draws inspiration from Nisbett and Miyamoto’s research on holistic versus analytic perception across cultures (2005), which showed that East Asian participants tend toward context-dependent, relational categorization while Western participants favor rule-based, feature-focused categorization. When presented with identical visual flower categorization tasks, two leading LLMs produced strikingly divergent answers that parallel human cultural differences documented in cross-cultural psychology research.

These results reveal that LLMs, though trained on similar internet-scale datasets, develop distinct perceptual and reasoning tendencies mirroring human cultural differences. Notably, Anthropic’s Claude model exhibited a more holistic, relationship-based perspective by grouping flowers based on visual pattern similarities, while OpenAI’s ChatGPT showed a more analytic, feature-focused approach by emphasizing specific petal characteristics.

Such observations raise critical questions for educational technology: When students interact with different LLMs, they may receive culturally distinct perspectives without awareness of embedded frameworks. As educational institutions increasingly implement AI assistants for tutoring, assessment, and content creation, educators must understand that AI tools are not culturally neutral, but embody specific cognitive orientations.

Our work advances the developing area of AI literacy by pinpointing ways educators can help students evaluate AI-generated content. Approaches include comparing responses from multiple AI systems, identifying cultural reasoning patterns, discussing embedded assumptions, seeking alternative perspectives, and reflecting on how varying cultural frameworks shape knowledge interpretation. Through such methods, students can develop metacognitive awareness about how AI systems encode cultural perspectives.

Educational implications extend beyond simple bias detection to fundamental questions about knowledge construction: How should educators present different cultural perspectives on knowledge organization? How might exposure to diverse AI reasoning patterns influence students’ cognitive development? Rather than viewing LLM differences as problematic, educators can leverage these variations to expose students to diverse reasoning patterns and enhance cross-cultural cognitive flexibility.

Keywords: Educational technology, large language models, AI literacy, cross-cultural perception, AI bias.