Whose assumptions a model treats as default. Trained predominantly on English-language internet text, models carry a particular set of norms about politeness, family, work, religion and law, and apply them as though universal. It shows up in small ways constantly and in significant ways occasionally.
Technical view
Measured by value survey instruments applied to models, showing alignment with particular national and demographic profiles, and by task performance on culturally specific knowledge. Sources include training data composition and preference annotation by a narrow annotator population. Mitigations include region-specific fine-tuning, culturally diverse preference data, and explicit steerability of norms, with the open question of who decides the default when no neutral option exists.
Related
Part of AI Terminology Index, section: Society and Impact.