GPT-5.2 and Gemini show divergent bias patterns across English and Swahili

The critical finding isn't that bias exists across languages, but that safety mechanisms (refusals, guardrails) actively diverge by language while bias itself shifts unpredictably. This means a model can be more permissive in one language while appearing aligned in another, creating a false sense of safety in non-English deployments.

This connects directly to the multilingual planning failures documented in early August, where low-resource languages systematically accumulated errors during conversion. But this bias study reveals a sharper problem: the safety layer itself becomes language-dependent. Unlike the TreeProbe benchmark (August 1st), which exposed gaps in training data representation, this work shows that alignment training itself fails to generalize across linguistic boundaries. The refusal asymmetry in GPT-5.2 mirrors the sycophancy vulnerabilities found in medical contexts, where context and framing determine whether guardrails activate, except here the context is simply the language of the prompt.

If OpenAI and Google release multilingual safety audits for their next model versions within the next two quarters that explicitly measure refusal rates and bias prevalence by language, that signals they're treating this as a priority. If they don't, and deployments in non-English markets proceed without language-specific safety benchmarks, that confirms the gap will persist in production systems.