●researcherThis reveals that identity-based prompting is an insufficient method for studying or mitigating cultural bias in LLMs.
●policyYou should account for implicit linguistic triggers when evaluating model alignment for diverse global populations.