Aligning Language Models to Local Cultural and Legal Norms

185 concepts98 papers1 proposal

When we align a large language model (LLM) to be "safe," we typically train it to refuse instructions for building weapons, generating hate speech, or leaking personal data. This paradigm treats safety as a universal constant. However, a growing body of research demonstrates that safety is deeply contingent on local legal frameworks, regional histories, and cultural taboos. What is benign in San Francisco might be politically inflammatory in Almaty, legally non-compliant in Beijing, or physically dangerous for a user in a jurisdiction where their identity is criminalized.

The Illusion of Universal Safety

Standard alignment techniques—such as Reinforcement Learning from Human Feedback (RLHF) or Direct Preference Optimization (DPO) trained on English-centric datasets—create an illusion of global safety. As Da Yin et al. demonstrate in SafeWorld: Geo-Diverse Safety Alignment, models optimized for Western norms frequently experience severe performance degradation when evaluated against non-Western legal and cultural realities.

This gap occurs because universal safety filters focus on high-level, globally recognized harms while remaining blind to localized social hierarchies and regional prejudices. For example, a model that successfully blocks Western-centric racism can completely fail to recognize or mitigate deep-seated regional prejudices, such as caste discrimination in India (Khyati Khandelwal et al. in Indian-BhED: A Dataset for Measuring India-Centric Biases in Large Language Models) or regionalism and academic background bias in South Korea (Hwaran Lee et al. in KoSBI: A Dataset for Mitigating Social Bias Risks Towards Safer Large Language Model Applications).

Why Simple Translation Fails

A common baseline approach to multilingual safety is translating English safety benchmarks into target languages. However, the authors of All Languages Matter: On the Multilingual Safety of LLMs show that direct translation fails to capture how definitions of harm shift across cultural boundaries. Translating a prompt preserves its literal meaning but strips away the localized socio-political context.

As Yasmine Ashraf et al. point out in Arabic Dataset for LLM Safeguard Evaluation, a completely harmless inquiry in a Western context can act as a sensitive political or religious trigger within the Arabic-speaking world.

Conversely, forcing Western-centric safety assumptions onto global users can lead to dangerous over-generalizations. In The Colorful Future of LLMs: Evaluating and Improving LLMs as Emotional Supporters for Queer Youth, Nitay Calderon et al. reveal that models aligned with Western values often provide advice assuming a universal baseline of safety for marginalized groups—such as encouraging public self-disclosure—which could be life-threatening in regions where those identities are legally criminalized.

The Pre-Training Bottleneck

Why can't we simply use system prompts to instruct a model to "be safe according to local culture"? Panuthep Tasawong et al. uncover a fundamental mechanism behind this failure in SEA-SafeguardBench: Evaluating AI Safety in SEA Languages and Cultures: instruction-based cultural prompting is largely ineffective if the underlying model was not sufficiently exposed to relevant regional data during its pre-training phase.

[System Prompt: "Follow local cultural norms"] 
               │
               ▼
┌────────────────────────────────────────┐
│       Large Language Model             │
├────────────────────────────────────────┤
│  Pre-training Data:                    │
│  ▲ Western/English Corpora (95%)       │  ► Fails to operationalize prompt:
│  ▼ Regional Corpora (5%)               │    Model lacks semantic scaffolding
└────────────────────────────────────────┘    for local taboos and laws.

Without a foundational representation of local taboos, historical narratives, and legal boundaries embedded in its weights, the model lacks the semantic scaffolding required to operationalize localized safety instructions. Safety alignment cannot be purely retrofitted at the surface level; it depends heavily on the cultural knowledge encoded during pre-training.

Operationalizing Localized Alignment

To move past these limitations, researchers are shifting from universal guardrails to localized alignment strategies that treat regional context as a primary variable.

Decoupling Intent from Culture

Instead of rewriting safety prompts from scratch for every region, the CAGE framework proposed by Chaeyun Kim et al. systematically separates an adversarial prompt's underlying intent (e.g., finding a legal loophole) from its cultural content (e.g., a specific local regulation). This allows safety teams to stress-test models against highly localized vulnerabilities without losing the structural rigor of global red-teaming methodologies.

The Local-to-Global Safety Synergy

Intuitively, one might worry that fine-tuning a model on highly specific regional harms would degrade its general capabilities or global safety boundaries. However, Aakanksha et al. demonstrate a surprising synergy in The Multilingual Alignment Prism: Aligning Global and Local Preferences to Reduce Harm.

Training models on culturally specific, localized harms provides a richer, more nuanced signal for learning the abstract concept of "harmfulness" than training on universal taboos alone. Exposure to local nuances actually hardens the model's broader safety boundaries, reducing the generation of globally prohibited content without imposing a heavy utility tax.

Key Foundational Work

AART: AI-Assisted Red-Teaming with Diverse Data Generation for New LLM-powered Applications (2023) — contributed 3 concept elements to this direction.
Indian-BhED: A Dataset for Measuring India-Centric Biases in Large Language Models (2023) — contributed 2 concept elements to this direction.
KoSBI: A Dataset for Mitigating Social Bias Risks Towards Safer Large Language Model Applications (2023) — contributed 2 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 91 papers
  2. Version 1 · 2026-10-04 · 114 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_localized_safety_alignment_3cb8c189,
  title = {localized safety alignment},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/localized-safety-alignment/rdr_3cb8c189c91b474b836722f591038795},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_localized_safety_alignment_3cb8c189_v2,
  title = {localized safety alignment},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/localized-safety-alignment/rdr_3cb8c189c91b474b836722f591038795/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake