Emergent Misinformation Genesis in Multi-Agent LLM Clinical Pipelines

Abstract

Multi-agent large language model (LLM) pipelines are deployed for clinical decision support under the assumption that collaboration improves safety. This paper shows that assumption is wrong. Multi-agent clinical pipelines spontaneously generate dangerous clinical assertions that no individual agent produces alone, with zero adversarial input. We term this Emergent Misinformation Genesis (EMG). We introduce the Emergent Misinformation Rate (EMR) metric with a three-way decomposition and evaluate across 4,800 trials with four model families and roughly 97,000 API calls. Two independent judges rate 70 to 87 percent of emergent assertions as clinically dangerous. The study identifies 74 critical drug interaction events, demonstrates collective delusion where individual agents reject 70 to 90 percent of the network's own assertions, and validates findings on 50 MIMIC-IV discharge summaries. All code, data, and 400 benchmark vignettes are released openly.

Ähnliche Arbeiten