When we prompt a large language model (LLM) with a seemingly neutral question, we expect a neutral answer. However, neutrality in machine learning is an illusion. Because LLMs are trained on internet-scale text corpora disproportionately generated by users in Western, Educated, Industrialized, Rich, and Democratic (WEIRD) societies, their internal statistical "center of mass" is heavily skewed.
When a model encounters ambiguity, it does not remain neutral; instead, it falls back on a systemic Western default. This phenomenon goes beyond a simple lack of factual knowledge—it is an architectural and statistical anchoring that projects Western cultural, psychological, and moral frameworks onto a global user base.
The Anatomy of the Default
The Western default is most visible when prompts lack explicit geographic or cultural context. In these ambiguous spaces, the model fills in the blanks using the highest-frequency cultural patterns from its training data.
This manifests across multiple mundane domains:
- Culinary and Social Customs: As Palta demonstrates using the FORK dataset, when models reason about everyday tasks like the order in which dishes are served, they consistently prioritize US-centric norms, even when they possess the latent knowledge to identify non-Western customs in more explicit contexts.
- Measurement Systems: When processing financial or physical units, models show a strong preference for Western standards like the US Dollar or the metric system. Bui et al. show that this bias persists regardless of the query's actual cultural context, degrading factual accuracy when regional units are required.
- The Illusion of Neutrality: What appears to be an unbiased baseline response is often a proxy for Western perspectives. Acquaye et al. show that when presented with culturally ambiguous prompts, models treat Western viewpoints as the standard, effectively marginalizing alternative global perspectives.
Linguistic Masking vs. Cultural Alignment
A common point of confusion is the assumption that a model's ability to speak a language implies an understanding of the accompanying culture. Recent work reveals that multilingual models are frequently just Western-aligned models wearing a linguistic mask.
[Multilingual Input] ──> [Shared Representation Space] ──> [Western Default Output]
(e.g., Farsi, (Erases cultural nuance; (Projected WEIRD norms
Hindi, Swahili) maps to dominant priors) & individualistic values)
When aligning multiple languages into a single representation space, the distinct cultural nuances of low-resource or non-Western linguistic groups are often erased. Havaldar et al. demonstrate this in the domain of emotion: multilingual models consistently project Western psychological frameworks onto Eastern cultures, favoring individualistic emotional responses (such as pride) in scenarios where collective or regulated emotions are culturally expected.
Furthermore, this misalignment creates a "fragility gap" in instruction following. Li et al. find that a model's ability to follow complex, multi-constraint instructions degrades when the prompt contains cultural references unfamiliar to the dominant training corpus. Instruction-following capability is not purely linguistic; it is deeply intertwined with the cultural context embedded in the training data.
Behavioral Failure Modes: Hallucination and Erasure
When forced to operate within non-Western contexts, the statistical pressure of the Western default causes models to fail in two distinct ways: cultural hallucination and homogenization.
Cultural Hallucination
When models lack sufficient representative data for a region, they do not gracefully admit ignorance. Instead, they project Western norms onto the target culture. Mukherjee et al. document how this leads to the hallucination of Western social standards in non-Western contexts, erasing local nuances. Similarly, Ramezani et al. show that models systematically misrepresent non-Western moral landscapes by overestimating the acceptance of controversial behaviors, failing to reflect the actual empirical beliefs of the populations they are modeling.
Homogenization and Stereotyping
When faced with specific regional artifacts, models often substitute them with high-frequency, media-popularized equivalents. Seth demonstrates this erasure using the DOSA dataset, showing that LLMs frequently misidentify localized Indian subcultural artifacts, defaulting instead to mainstream or media-amplified stereotypes.
Why Simple Fixes Fail
If the bias stems from data distribution, can we simply prompt the model to change its perspective? Evidence suggests that the Western default is deeply embedded and highly resistant to superficial interventions.
Contextual Reversion
Lee et al. show that even when explicitly instructed to adopt a specific regional viewpoint (e.g., evaluating hate speech from a non-North American perspective), LLMs frequently default back to the cultural norms of their primary training data. For smaller models, this reversion is even more pronounced. Rao et al. note that when models are prompted to follow non-Western ethical policies, they often struggle with dilemmas that prioritize communal values over individualistic ones, actively generating justifications that attempt to twist the scenario back into an individualistic framework.
The Computational Tax
When test-time prompting techniques like Chain-of-Thought (CoT) do succeed in forcing the model to overcome its default and reason accurately about underrepresented cultures, they introduce a hidden equity issue. Bui et al. discover that overcoming these default biases requires significantly more test-time compute—up to a 300% increase in token generation costs.
This imposes a literal computational tax on users from non-Western backgrounds, who must expend more resources and time simply to get the model to bypass its Western defaults and generate factually accurate, culturally grounded responses.