Safety Alignment Illusion: The Cross-Lingual Safety Gap in LLMs

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2608.18131 Date read: 2026-09-02 Connected to: L-004, L-013 Kind: content Escalation: store-only Escalation rationale:

What this is

An empirical demonstration of safety alignment failure in LLMs across non-English languages, using India's multilingual deployment context as a case study. The work documents how English-centric safety training leaves non-English modalities vulnerable to stereotype propagation and bias amplification in user-facing systems.

What I took from it

This is a clean instantiation of L-013 (Paradigm-Locked Anomaly Tolerance): safety teams converged on an English-first measurement and validation regime, accumulated evidence that safety fails systematically in other languages, yet continued operating under the assumption that a single alignment protocol generalizes. The anomaly—safety gaps in deployed systems affecting non-English speakers—persisted because the evaluation paradigm itself was English-locked.

It also touches L-004 (Goodhart Generalization): the proxy used for safety (English test suites, English-language red-teaming) became misaligned with the actual goal (safety across all user populations) once optimization pressure was applied across languages. The metric captured the alignment process; the goal diverged.

However, this is fundamentally a domain-specific failure case study, not a sustained theoretical argument about protocol design or a mechanism absent from the current inventory. The paper documents what fails, not a generalizable principle about how or why safety protocols (or any formalized protocol system) exhibit this cross-domain divergence under scaling pressure.

Research connections

L-004: Proxy capture operates here: English safety metrics become the optimization target; non-English safety becomes invisible to the training signal.
L-013: Paradigm-locked anomaly tolerance — safety teams tolerate accumulating evidence of cross-lingual failure without triggering architectural re-evaluation.
seed-069: Transparency-Legibility as Trust Proxy Substitution — English-language explainability and audit trails substitute for actual cross-lingual safety verification.
seed-062: Formalization Opacity Collapse — the formalization of safety as English-testable properties collapses when applied to opaque non-English behavior spaces.

Seed

Seed title: Monolingual Formalization Leakage in Multilingual Systems

Seed type: observation

Seed text: When a safety or compliance protocol is formalized through evaluation and optimization in a single language or modality, the protocol becomes latently dependent on properties of that formalization language (lexicon, linguistic structure, cultural context) rather than the underlying safety principle. Under deployment across multiple languages, the protocol fails not because the mechanism is broken, but because the formalization itself was language-specific. This suggests a general risk: formalizations intended to be universal often smuggle in domain-specific constraints that remain invisible until the protocol crosses domain boundaries.