Mozilla researchers combined two projects to test how AI guardrails perform across languages in humanitarian contexts. Using Mozilla.ai's any-guardrail framework with 60 real-world asylum seeker scenarios in English and Farsi, they evaluated whether guardrails maintain consistent safety standards when processing non-English inputs.

Daniel Nissani's any-guardrail provides customizable interfaces for classifier-based and generative guardrails, while Roya Pakzad's Multilingual AI Safety Evaluations contributed the humanitarian case study. The team tested three guardrails - FlowJudge, Glider, and AnyLLM (GPT-5-nano) - against policies developed in both English and Farsi, grounded in the Multilingual Humanitarian Response Eval dataset covering six dimensions including safety, accuracy, and non-discrimination.

The experiment revealed significant discrepancies when guardrails evaluated Farsi versus English responses, with differences of 2+ points on 5-point scales. This matters because humanitarian contexts require nuanced understanding of socio-political factors like sanctions, displacement drivers, and regional variations. When an asylum seeker asks about cryptocurrency regulations for education funding under sanctions, purely linguistic competence proves insufficient - guardrails need domain-specific knowledge to provide safe, accurate guidance.