Refusal safety transfers less reliably for some low-resourced languages
Is Refusal Direction Universal across Low-Resourced Languages?
Computation and LanguageMachine Learning
Summary
This paper looks at how well AI systems can refuse unsafe requests in different low-resourced languages. Previous work found refusal instructions transfer almost perfectly across well-resourced, safety-aligned languages, but the authors show that for some low-resourced languages that are very different in behavior, the transfer is weak or fails. They tested three new low-resourced languages and found the effect is not simply universal or not, but varies depending on similarity and alignment. This means some languages might face weaker safety protections when using AI.
What this means in practice
- •For ai safety engineers: Adjust safety protocols for AI systems based on language resource levels to ensure consistent refusal behavior in diverse languages.
- •For multilingual ai developers: Design multilingual models that consider language alignment differences to improve refusal and safe response behaviors in low-resource languages.
Authors
Anastasiia Ivanova
Abstract
Refusal directions have been shown to transfer cross-lingually with near-perfect effectiveness across safety-aligned languages (Wang et al., 2025), reportedly regardless of resource level. However, the only truly low-resourced language evaluated is Yoruba, which is strongly misaligned and shows limited transfer, leaving open how refusal directions behave for other low-resourced languages. We extend PolyRefuse to three additional low-resourced languages (Belarusian, Bashkir, and Tajik) and replicate the pipeline of (Wang et al., 2025) on Qwen2.5-14B-Instruct. We find that universality is graded rather than binary: directions from safety-aligned and partially-aligned languages transfer broadly, while those from strongly misaligned languages fail to reliably modulate refusal behavior on any target tested, including other misaligned languages, once ceiling effects are accounted for. Geometric alignment and behavioral transfer are further shown to be partially dissociated. These findings complicate the universality claim and suggest that users of the least-resourced, most strongly misaligned languages may face weaker safety guarantees.