A randomized controlled trial with 1,298 participants revealed inconsistent performance when laypeople used large language models for medical decision-making assistance. The study measured diagnostic accuracy and treatment recommendations across various health scenarios, finding that AI assistance improved outcomes in some cases while potentially hindering performance in others. This mixed efficacy represents a critical juncture for consumer health AI applications. Unlike previous studies that focused primarily on healthcare professionals, this research examined real-world scenarios where ordinary individuals might rely on AI for health guidance. The findings challenge the assumption that more information access automatically leads to better health decisions. For health-conscious adults, the implications are significant—AI tools may excel at pattern recognition for certain conditions but could introduce overconfidence or misinterpretation in others. The variability suggests that AI medical assistance requires careful implementation with clear limitations rather than broad deployment. This controlled evidence provides the first rigorous assessment of how AI actually performs when used by the general public for health decisions, highlighting the need for nuanced approaches to consumer health AI rather than blanket adoption or rejection.