Majority Poisoned Documents Induce Conformity Bias in RAG
November 2, 2026
Research shows that increasing the number of mutually agreeing poisoned documents in RAG contexts significantly reduces misinformation detection. For Gemma-3-12b, detection rates dropped from 25% with one poisoned document to 10% when six documents agreed on the false claim.
HOW THIS AFFECTS YOU
●
builderYou must account for majority-bias vulnerabilities when designing retrieval-augmented generation pipelines.
●
researcherThis provides evidence of conformity-like responses to misinformation within LLM retrieval architectures.