Safety Systems Fail by Targeting Lexical Form Over Behaviour
October 23, 2026
Research shows automated moderation systems disproportionately flag African American English and specific marginalized identities due to a focus on vocabulary rather than intent. The authors argue that safety protocols should evaluate harmful behaviors, such as isolation or secrecy demands, which transfer across languages more reliably than surface-level words.
HOW THIS AFFECTS YOU
●
researcherThis highlights the need for cross-lingual behavioral benchmarks rather than lexical toxicity lists.
●
policyYou should shift evaluation frameworks from keyword-based moderation to behavioral-based detection to reduce bias.