Mind the Gesture: Evaluating AI Sensitivity to Culturally Offensive Non-Verbal Gestures

AI systems show significant US-centric bias: Text-to-image systems perform better at detecting offensive gestures in US contexts than in non-US ones, suggesting that AI training data and design prioritizes American cultural norms over global perspectives.

Large language models tend to over-flag gestures as offensive: LLMs demonstrate a pattern of incorrectly labeling gestures as inappropriate more often than warranted, which could lead to unnecessary censorship or misunderstandings in cross-cultural communication.

Vision-language models default to US interpretations of universal concepts: When asked about culturally neutral actions like wishing someone luck, VLMs frequently suggest gestures that are culturally inappropriate in many parts of the world, revealing a fundamental gap in cultural awareness.

The scale of the problem is substantial and systematic: The research evaluated 288 gesture-country pairs across 25 gestures and 85 countries, demonstrating that cultural insensitivity in AI is not isolated incidents but rather a widespread issue affecting multiple AI system types.

Culturally-aware safety mechanisms are urgently needed: As AI systems become more globally deployed, the paper emphasizes that current safeguards are insufficient and new approaches are required to ensure AI systems respect and accurately interpret cultural differences rather than perpetuating Western-centric biases.