Multilingual LLM outputs can vary across sociocultural contexts. However, evidence of cultural grounding can be misleading: identity labels may be inferred from explicit or indirect textual cues, while names and wording can reveal the source language. Treating all these signals as evidence of cultural grounding may obscure potential biases. We present a human-validated, multi-agent audit that separates three questions: whether outputs reproduce social biases, whether identity groups are represented differently, and whether outputs reflect cross-cultural patterns. The study analyzes 89,253 outputs from 12 LLMs in English, French, and Chinese, spanning 18 occupations and three task conditions. We find that bias representation varies systematically across languages and tasks. Removing direct identity cues sharply reduces identity-label prediction in English and Chinese, but has a much smaller effect in French. Across all language-genre settings, the cultural context associated with the source language receives the highest average relevance score, with moderate agreement between automated and human ratings. However, the ability to identify the source language drops substantially after translation and again after masking names. Without these controls, multilingual audits may mistake surface cues for cultural understanding, leading to misleading conclusions about cross-cultural variation and bias. Our audit offers a practical framework for separating such shortcuts from more meaningful cross-cultural patterns.
Figures & tables
Figure 1: Overview of the experiment setup.
Figure 2: task-level Balance and Stereotype scores by language. Each point corresponds to one model in one task condition. Shaded ellipses summarize dispersion across models within each condition.
Language
Genre
Full text
Cue-deleted
Δ AUC
EN
Story
.999 [.996, 1.000]
.634 [.555, .713]
−.365 [ −.444 , −.286 ]
EN
News
1.000 [1.000, 1.000]
.771 [.703, .827]
−.230 [ −.297 , −.173 ]
FR
Story
.998 [.992, 1.000]
.994 [.988, .999]
−.003 [ −.010 , .002]
FR
News
.997 [.989, 1.000]
.960 [.935, .981]
−.037 [ −.060 , −.018 ]
ZH
Story
.994 [.984, 1.000]
.674 [.601, .744]
−.320 [ −.390 , −.252 ]
ZH
News
.988 [.976, .997]
.646 [.573, .717]
−.342 [ −.411 , −.273 ]
Table 1: Identity-label probe evaluated on held-out model–occupation groups in the paired subset (200 narratives per cell). Full text is the positive-control condition; values are ROC AUC with 95% intervals computed by resampling these groups, and Δ is cue-deleted minus full-text AUC.
Figure 3: Shared UMAP projection of the full-text multilingual narrative embeddings. Pink points and solid contours denote narratives associated with females, and blue points and dashed contours denote narratives associated with males; squares indicate Story and circles indicate News . Contours indicate density regions (20/50/80 percentiles). Display-only percentile trimming is described in the methodology; the UMAP axes are not used for inference.
Figure 5: Cross-cultural relevance pattern among marker-bearing narratives. Each cell shows narrative-weighted mean automated relevance on the original 1–7 scale by source language, evaluated language-linked reference context, and genre.
As Large Language Models (LLMs) develop stronger multilingual capabilities, their sensitivity to culturally diverse entities becomes increasingly important. Prior work by Naous et al. (2024) has shown that LLMs often favor Western-associated entities in Arabic. Due to the lack of entity-centric multilingual benchmarks, it remains unclear if such biases also manifest in various non-Western languages. In this paper, we introduce Camellia, a benchmark for evaluating entity-centric cultural biases in nine Asian languages, spanning six Asian cultures. Camellia includes 19,530 manually annotated entities associated with the covered Asian or Western cultures, as well as 2,173 masked contexts for these entities derived from social media posts. Using Camellia, we evaluate cultural biases in four recent multilingual LLMs across three tasks: cultural context adaptation, sentiment association, and entity extractive QA. Our analyses show that LLMs struggle with cultural adaptation across these languages, with performance differing across models developed in different regions. We further observe that different LLM families can hold distinct biases, reflected in the ways they link cultures to particular sentiments. Lastly, we find that LLMs can struggle with context understanding in some Asian languages, creating performance gaps between cultures in entity extraction.
Tarek Naous, Anagha Savit, Carlos Rafael Catalan +17
Georgia Institute of Technology · Samsung R&D Institute Philippines · Sungkyunkwan University +6
Despite their impressive capabilities, multilingual large language models (MLLMs) frequently exhibit inconsistent behaviour when the prompt's language changes. While such adaptation is generally desirable, it becomes a critical failure when a user's identity is explicitly defined. For instance, given a fixed British persona and an ambiguous everyday knowledge query about literature, the prompt's language frequently overwrites the system persona -- yielding Shakespeare in English but Cervantes in Spanish. To robustly quantify this Cross-lingual Cultural Inconsistency, we introduce Singleton Fleiss's κS, a metric mathematically resilient to hallucinations. For mitigation, we propose Cross-lingual Cultural Consistent Preference Optimisation (C-3PO), a consensus-driven alignment framework. C-3PO achieves up to a 0.13-point absolute increase in κS over unaligned models, consistently outperforming strong prompting and representation steering baselines whilst preserving explicit user identities, cultural neutrality and intrinsic cultural knowledge. Empirical evaluations demonstrate this inconsistency disproportionately affects lower-resource languages like Indonesian and Persian. Finally, early decoding of intermediate layers reveals that MLLMs implicitly personalise outputs towards the prompt language's stereotypical culture as forward-pass representations stabilise.
Lucas Resck, Isabelle Augenstein, Anna Korhonen
Language Technology Lab, University of Cambridge · University of Copenhagen
Multilingual fluency often invites a stronger assumption: a model that can speak a user's language must also understand the culture encoded by that language. We call this the Illusion of Cultural Alignment. To test this assumption directly, we introduce MSQA, a benchmark of 1,064 natively sourced questions across 11 language groups, five cultural dimensions, and three difficulty tiers. Unlike translated benchmarks, MSQA targets locally grounded knowledge and reduces shortcuts from English-centric cross-lingual transfer. Evaluating 18 LLMs, we find substantial cultural degradation and a pronounced Locality Effect: cultural competence tracks pre-training exposure more closely than general reasoning ability. We further show that common inference-time remedies do not dissolve the illusion. Models remain overconfident on unfamiliar cultural questions, repeated sampling yields unstable rather than reliable correctness, and retrieval augmentation helps unevenly on long-tail facts. These findings indicate that cultural alignment cannot be inferred from multilingual ability alone and requires deeper intervention than calibration, sampling, or retrieval at inference time