Key points are not available for this paper at this time.
Recent years have seen substantial investments in AI-based tools designed to detect offensive language at scale, aiming to moderate social media platforms, and ensure safety of conversational AI technologies such as ChatGPT and Bard. These efforts largely treat this task as a technical endeavor, relying on data annotated for offensiveness by a global crowd workforce, without considering crowd workers' socio-cultural backgrounds or the values their perceptions reflect. Existing research that examines systematic variations in annotators' judgments often reduces these differences to socio-demographic categories along racial, or gender dimensions, overlooking the diversity of perspectives within such groups. On the other hand, social psychology literature highlights the crucial role that both cultural and psychological factors play in human perceptions and judgments. Through a large-scale cross-cultural study of 4309 participants from 21 countries across eight cultural regions, we demonstrate substantial cross-cultural and individual moral value-based differences in interpretations of offensiveness. Our study reveals specific regions that are significantly more sensitive to offensive language. Furthermore, using the Moral Foundations Theory, we study the underlying moral values that contribute to these cross-cultural differences. Notably, we find that participants' moral values play a far more important role in shaping their perceptions of offensiveness than geo-cultural distinctions. Our investigation, using a non-monolithic framework to understand cross-cultural moral concerns, reveals crucial insights that can be extrapolated to building AI models for the pluralistic world. Our results call for more extensive consideration of diverse human moral values when deploying AI models across diverse geo-cultural contexts.
Davani et al. (Mon,) studied this question.