African Ethnoveterinary Terminology Creates Blind Spots in AI Biosecurity Guardrails : Preliminary Findings by fatika

Afri­can Eth­novet­eri­nary Ter­minol­ogy Creates Blind Spots in AI Biose­cu­rity Guardrails : Pre­limi­nary FindingsAI biose­cu­rity guardrails may be sys­tem­at­i­cally blind to lo­cal dis­ease lan­guage and it may mat­ter most in the places least rep­re­sented in AI safety train­ing data.I had ear­lier built a bench­mark eval­u­at­ing fron­tier model perfor­mance on Afri­can live­stock and eth­novet­eri­nary knowl­edge sys­tems. Llama 3.1 8B achieved only 43% over­all ac­cu­racy, though it pe...