Search

LLM Safety Filters Fail Differently Across Dialects and Explicit Identity

Research shows language models refuse requests more often when users state their identity explicitly, but bypass safety guardrails when using dialect signals like AAVE.