Presumed Cultural Identity: How Names Shape LLM Responses

TL;DR

Presumed Cultural Identity investigates how user names influence LLM personalization by eliciting cultural presumptions in outputs. The authors assemble a cross-cultural test bed using 900 names across 30 cultures, a CANDLE knowledge graph for concrete cultural assertions, and evaluations over 4 open models plus a closed model to quantify name-driven biases. They compare LLM-as-judge and assertion-entailment approaches, validate with human annotations, and perform facet- and model-specific analyses, revealing robust biases that vary by culture, model, and queried facet (notably clothing and tradition). The work highlights the risks of name-based personalization—such as cultural flattening and underrepresented groups being misrepresented—and outlines design considerations for more nuanced, transparent, and fair personalization in AI systems.

Abstract

Names are deeply tied to human identity. They can serve as markers of individuality, cultural heritage, and personal history. However, using names as a core indicator of identity can lead to over-simplification of complex identities. When interacting with LLMs, user names are an important point of information for personalisation. Names can enter chatbot conversations through direct user input (requested by chatbots), as part of task contexts such as CV reviews, or as built-in memory features that store user information for personalisation. We study biases associated with names by measuring cultural presumptions in the responses generated by LLMs when presented with common suggestion-seeking queries, which might involve making assumptions about the user. Our analyses demonstrate strong assumptions about cultural identity associated with names present in LLM generations across multiple cultures. Our work has implications for designing more nuanced personalisation systems that avoid reinforcing stereotypes while maintaining meaningful customisation.

Figures (13)

• Figure 1: Example of an interaction with an LLM with an identity presumption based on the name
• Figure 2: Experimental Setup
• Figure 3: Default Bias values averaged over Models and Facets. For details refer to \ref{['sec: bias_cal']}.
• Figure 4: Bias across models above the default bais. For calculation of bais refer to section \ref{['sec: bias_cal']}
• Figure 5: Box plot showing comparison of bias for countries values (averaged over 4 models) for each facet.