Source-linked AI summary
Bridging the Multilingual Safety Divide: Efficient, Culturally-Aware Alignment for Global South Languages
Somnath Banerjee, Rima Hazra, Animesh Mukherjee
TL;DR
Global South deployment exposes a gap between English-centric safety assumptions and real use involving low-resource languages, code-mixing, and cultural variation. The paper synthesizes evidence on multilingual safety failures and proposes efficient, culturally grounded, participatory alignment; the findings show that targeted defenses can address these failures without prohibitive compute.
Problem
English-centric benchmarks, safety pipelines, and edits do not reliably cover low-resource, code-mixed, culturally specific, or cross-language settings.
Method
The paper synthesizes recent empirical work and proposes resource-aware evaluation, parameter-efficient steering, culturally grounded preference data, participatory alignment, and multilingual auditing.
Results
Safety weakens on low-resource and code-mixed inputs, cultural harms persist beyond toxicity scores, and English edits often fail to transfer; targeted defenses improve multilingual safety.
Takeaways & Limitations
Multilingual safety should be treated as a core, culturally grounded, and participatory requirement for equitable AI in underrepresented regions.
Abstract
from arXiv · showhide
Large language models (LLMs) are being deployed across the Global South, where everyday use involves low-resource languages, code-mixing, and culturally specific norms. Yet safety pipelines, benchmarks, and alignment still largely target English and a handful of high-resource languages, implicitly assuming safety and factuality ''transfer'' across languages. Evidence increasingly shows they do not. We synthesize recent findings indicating that (i) safety guardrails weaken sharply on low-resource and code-mixed inputs, (ii) culturally harmful behavior can persist even when standard toxicity scores look acceptable, and (iii) English-only knowledge edits and safety patches often fail to carry over to low-resource languages. In response, we outline a practical agenda for researchers and students in the Global South: parameter-efficient safety steering, culturally grounded evaluation and preference data, and participatory workflows that empower local communities to define and mitigate harm. Our aim is to make multilingual safety a core requirement-not an add-on-for equitable AI in underrepresented regions.
Introduction and Motivation
Global South AI deployment occurs under limited compute, scarce labeled data, and multilingual conditions that English-centric safety pipelines do not represent. This work synthesizes evidence on multilingual benchmarks, cultural harm, code-mixed failures, and multilingual model editing to develop a resource-aware agenda.
- Global South users encounter low-resource languages, code-mixed text, and culturally situated discourse that differ from high-resource safety benchmarks.These conditions include discussions of migration, religion, politics, and local institutions.
- Multilingual safety is framed as a fairness and participation issue, not merely a technical challenge.
- The synthesis covers multilingual safety benchmarks, parameter-efficient steering, cultural harm evaluation, code-mixed failures, and multilingual model editing.It distills these findings into a resource-aware blueprint for Global South students and early-career researchers.
Evidence of a Multilingual Safety Divide
Recent evidence shows that safety and factuality do not reliably transfer from English to low-resource, non-Latin, code-mixed, or culturally specific settings. The reviewed studies identify targeted steering, culturally grounded preferences, and multilingual auditing as responses to these failures.
- Low-resource and non-Latin inputs produce more unsafe or under-moderated outputs, while English-only finetuning and translation filtering miss risks or misclassify content.
- Updating roughly 3% of language-specific functional-head parameters improves safety across all 10 languages while preserving MMLU and TruthfulQA capability.The method targets attention heads associated with harmful behavior, avoiding unrealistic full retraining.
- Models judged safe by standard toxicity metrics can still generate outputs that local annotators consider culturally insensitive or harmful.Culturally grounded preference data substantially lowers such responses without sacrificing, and sometimes improving, substantive question answering.
- Code-mixed harmful-request attack success rises from ∼9% in monolingual English to ∼69% in code-mixed paraphrases, exceeding 90% in some Arabic and Hindi cases.These vulnerabilities also occur in natural user-generated code-mixed inputs.
- English knowledge edits often fail to transfer to low-resource languages, with factual consistency dropping outside English even when English accuracy improves.Model merging reduces but does not eliminate these multilingual gaps.
A Resource-Aware Blueprint
The proposed blueprint addresses multilingual safety under limited compute through local evaluation, efficient steering, participatory alignment, and cross-language auditing. It combines targeted parameter updates, attribution-guided corrections, culturally grounded preferences, and verification beyond English.
- The agenda recommends multilingual and code-mixed testbeds, including XThreatBench, to detect region-specific safety failures.
- Efficient mitigation uses language-specific functional heads at ≈3% of parameters or attribution-guided corrections instead of full retraining.
- Participatory alignment replaces generic filters with culturally grounded preference data collected through community-driven annotation.
- Multilingual auditing verifies that safety patches and factual edits work across low-resource languages rather than becoming English-only upgrades.
Conclusion
The synthesis concludes that multilingual safety failures are systemic but solvable with resource-efficient, culturally grounded defenses. Equitable alignment requires replacing English-centric transfer assumptions with intrinsically multilingual and participatory strategies.
- Code-mixed queries, cultural nuances, and knowledge propagation expose systemic multilingual safety failures.
- Parameter-efficient steering and attribution-guided defenses can support robust multilingual alignment without prohibitive compute resources.
- Bridging the divide requires intrinsically multilingual, culturally grounded, and participatory alignment rather than English-centric transfer assumptions.