Language-Specific Gaps in AI Safety Training Datasets

Research indicates that claims of multilingual safety in large language models (LLMs) are often not supported by robust evidence at the individual language level. An audit of 21 resources across 25 language slices, covering low-, mid-, and high-resource languages (Hausa, Swahili, French), revealed recurring deficiencies. These include gaps in data provenance, annotation reliability, access, harm-taxonomy coverage, and data reuse, which inconsistently correlate with resource levels.

Why it matters

The findings challenge the perceived safety and reliability of AI systems for non-English speaking populations, highlighting a critical disparity in AI development and deployment. This can lead to increased risks for organizations operating globally, potentially undermining trust and adoption of AI technologies in diverse linguistic markets.

Key insights

• LLM providers' claims of multilingual safety, based on benchmarks, are often unreliable when examined at the individual language level.
• Auditing 21 resources and 25 language slices revealed significant gaps in AI safety training datasets for non-English languages.
• Deficiencies include issues with data provenance, the reliability of annotations, data accessibility, and comprehensive harm-taxonomy coverage.
• Problems with data reuse were also identified in multilingual safety training datasets.
• These identified gaps do not consistently track with the resource level of the language (e.g., low-resource Hausa vs. high-resource French), indicating systemic issues beyond simple resource scarcity.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.13695