Large language models (LLMs) are typically aligned using safety datasets and human feedback primarily sourced from English-speaking, Western contexts. While this process effectively suppresses globally recognized harms like "how to build a bomb," it creates a significant "safety gap" when these models are deployed globally. Research across dozens of regional contexts—from Southeast Asia to the Middle East—reveals that safety is not a universal constant but a culturally situated variable.
The Cross-Lingual Safety Gap
A recurring finding in recent research is that a model’s refusal behavior in English does not reliably transfer to other languages. Wenxuan Wang and colleagues demonstrated this in All Languages Matter, showing that models like ChatGPT often provide unsafe responses to the same harmful prompt when translated into lower-resource languages. This vulnerability is further quantified by Felix Friedrich through the M-ALERT benchmark, which uncovers consistent safety failures across French, German, Italian, and Spanish LLMs Lost in Translation.
The intuition behind this gap is that safety alignment (such as RLHF or DPO) acts as a "wrapper" around the model's base knowledge. If the alignment data is English-centric, the model learns to associate "harmfulness" with English linguistic patterns. When prompted in Kazakh or Arabic, the model may revert to its unaligned base state, failing to recognize the adversarial intent because the "safety guardrails" were never mapped to those specific linguistic tokens.
The Failure of Translation-Based Safety
Early attempts to fix this relied on machine-translating English safety datasets. However, researchers like Raviraj Joshi argue that literal translation fails because the concept of harm is deeply tied to social context CultureGuard. For example:
- Regional Scams: A safety prompt about a US-centric Social Security scam is irrelevant in a region where that institution doesn't exist.
- Legal Nuance: Advice that is safe in California—such as recommending self-disclosure to a marginalized youth—can be life-threatening in jurisdictions where certain identities are criminalized The Colorful Future of LLMs.
- Social Taboos: Definitions of what is "scary" for children or "offensive" in a religious context vary wildly across borders PG-Story, Arabic Dataset.
To address this, the field is moving toward transcreation—adapting the intent of a safety prompt to local cultural and legal realities rather than just translating the words.
Localized Benchmarking and Taxonomy
To measure these gaps, researchers are developing region-specific benchmarks that move beyond general toxicity. Jianyu Wang and the SeaLLMs team introduced SeaBench, which uses native exam questions and local scenarios to evaluate models in Southeast Asian contexts. Similarly, IndoSafety extends safety taxonomies to include region-specific sensitivities like local supernatural beliefs and national ideologies.
These benchmarks often reveal that even "region-centric" models—those trained on local data for better fluency—do not automatically inherit better safety. Without explicit safety training on local cultural nuances, these models remain as vulnerable to regional "jailbreaks" as their Western counterparts Qorǵau.
Mechanisms for Pluralistic Alignment
How can we align models for a global audience without a single "universal" rater? Several mechanisms have emerged:
- Reward Gap Optimization (MPO): Weixiang Zhao proposed using a model's own robust safety signals in English to calibrate its behavior in other languages, effectively "transferring" the safety boundary across the linguistic latent space MPO.
- Culturally Adaptive Red-Teaming (CAGE): Chaeyun Kim developed a framework that separates adversarial intent from cultural content, allowing researchers to systematically test for vulnerabilities in local legal loopholes or social conflicts CAGE.
- Synthetic Culturally-Aware Pipelines: Tools like CultureGuard use LLMs to "re-contextualize" safety data, swapping Western entities for local ones to create high-quality training data without the prohibitive cost of global human annotation CultureGuard.
Ultimately, as Arkadiy Saakyan notes, roughly 10% of safety content is "culturally sensitive"—meaning its harm is in the eye of the beholder Quantifying the Salience of Geo-Cultural Values. This suggests that the future of AI safety lies not in a single global filter, but in pluralistic frameworks that can adapt to the specific geo-cultural values of the user.