When we train machine learning models or conduct large-scale behavioral studies on web data, it is easy to fall into the trap of assuming that the data reflects a universal human standard. We treat large datasets as neutral mirrors of global behavior. However, data is never truly placeless; it is deeply anchored in specific geographies, languages, and cultural institutions.
The core insight of this research direction is that the vast majority of foundational datasets are heavily skewed toward the Anglosphere—specifically English-speaking users in Western countries like the United States and the United Kingdom. When models trained on this data are deployed globally, their generalizability breaks down. They do not fail because of algorithmic flaws, but because they are attempting to apply highly localized cultural norms, linguistic patterns, and social expectations to a diverse global population.
The Illusion of the Universal User
Why does this bias persist? Large-scale social media and web datasets often feel representative because of their sheer volume. Yet, as Emma S. Spiro et al. observe in Is anyone out there?: unpacking Q&A hashtags on twitter, social media research frequently relies on English-language corpora and Western user bases. This reliance creates an "imagined community" that masks how fundamentally social expectations, tagging behaviors, and communication styles vary across different cultures and linguistic groups.
When researchers assume a dataset is globally representative simply because it is large, they build systems with systemic blind spots. For instance, in behavioral and public health research, insights are frequently constrained by these exact boundaries. Lauren Sinnenberg et al. demonstrate in Content Analysis of Metaphors About Hypertension and Diabetes on Twitter: Exploratory Mixed-Methods Study that public health insights derived from social media are often restricted to English-language posts from specific regions, overlooking the unique conceptual frameworks and metaphors that diverse global populations use to describe chronic conditions.
How Localized Norms Distort Model Behavior
This geographic concentration becomes problematic when models are tasked with understanding nuanced social concepts, such as safety, toxicity, or identity. Social norms are not uniform; what is understood as harmful or standard in one region may be perceived entirely differently in another.
Social and Safety Concepts
When datasets are annotated by a geographically restricted group, their subjective definitions are baked into the model. In SemEval-2023 Task 10: Explainable Detection of Online Sexism, Hannah Rose Kirk et al. point out that while fine-grained sexism detection is vital, relying on annotations exclusively from women residing in the UK limits the dataset's applicability elsewhere, as definitions of sexism vary significantly across regions. Similarly, Svetlana Kiritchenko et al. show in Aporophobia: An Overlooked Type of Toxic Language Targeting the Poor that toxic language detection models are heavily constrained by English-language resources and US-centric platform demographics, making them ill-equipped to recognize socioeconomic biases in other cultural contexts.
Conceptual and Technical Priors
This bias extends beyond social commentary into basic conceptual and technical tasks. Vision-Language Models (VLMs) inherit narrow worldviews from their training corpora. Stella Frank et al. demonstrate in VISaGE: Understanding Visual Generics and Exceptions that conceptual datasets restricted to American English fail to capture how exceptions and visual norms are perceived or represented in other linguistic and cultural contexts.
Even fields assumed to be purely technical, such as software engineering, suffer from this localized framing. Chetan Bansal et al. in The Usage of Web Search for Software Engineering and Foyzul Hassan in An Empirical Study of Software Exceptions in the Field using Search Logs emphasize that developer troubleshooting behavior studies are overwhelmingly focused on English-speaking regions. This ignores how global developer communities interact with localized documentation, language-specific search patterns, and distinct cultural approaches to problem-solving.
The Downstream Consequences
When the datasets used to evaluate models are themselves biased, our metrics for success become untrustworthy. Percy Liang et al. address this in Trustworthy Social Bias Measurement, arguing that the reliance on English-language corpora and Western demographic data limits the generalizability of bias metrics themselves. A metric designed around Western social dynamics cannot accurately reflect normative ideals or social harms in non-Western environments.
This evaluation blind spot leads to active systemic failures in production models. A clear example is documented by Jieyu Zhao et al. in A Rose by Any Other Name would not Smell as Sweet: Social Bias in Names Mistranslation. The authors reveal that state-of-the-art machine translation systems frequently mistranslate or "Anglicize" personal names associated with racial and ethnic minorities. Because the underlying training datasets are derived from narrow national or urban contexts, they fail to represent global name diversity, causing translation systems to default to dominant Western naming conventions.
[Narrow Training Data] -> (US/UK/English Centric)
│
▼
[Biased Model Priors] -> (Assumes Western social norms, naming patterns, & metaphors)
│
▼
[Global Deployment] -> (Fails on regional dialects, localized terms, & cultural nuances)
Moving Beyond the Anglosphere
To build models that genuinely generalize, the research community must transition from treating English-centric data as a default universal proxy to intentionally curating localized, multilingual resources.
Several efforts highlight the necessity of this shift by focusing on specific regional contexts:
- Targeted Hate Speech Detection: Hasan Kerem Şeker et al. developed HATECAT-TR: A Hate Speech Span Detection and Categorization Dataset for Turkish to capture the nuanced, platform-specific hate speech patterns unique to Turkish social media, which generic global models fail to catch.
- Multilingual Domain Adaptation: Maxime Masson et al. demonstrate in Optimal strategies to perform multilingual analysis of social content for a novel dataset in the tourism domain how few-shot learning can be optimized to handle regional, multilingual social content effectively without requiring massive, centralized English corpora.
By recognizing that data is inherently situated within geographic and linguistic boundaries, researchers can move away from the false assumption of universal generalizability and toward systems designed to navigate a genuinely polycentric world.
Key Foundational Work
Is Anyone Out There? Unpacking Q&A Hashtags on Twitter (2014) — contributed 1 concept elements to this direction.
From Web Pages to Web Communities (2009) — contributed 1 concept elements to this direction.