The Inefficacy of Artificial Intelligence Large Language Models in Healthcare: A Clinical and Statistical Perspective

Objective: This perspective piece examines the role of Large Language Models (LLMs) in healthcare, arguing that despite significant investment, these models have had only a limited impact. Moreover, we argue that LLMs must replicate key phases of primary healthcare delivery to be a force multiplier, a necessary condition to address the global burden of disease. Discussion: We argue that LLMs lack the metacognitive capacity for ranked, dynamic reasoning. This is evidenced by clinically dangerous hallucinations and inability to perform unless complete information is provided. We extend clinical critiques with a statistical argument and a simulation exercise demonstrating that LLM-based diagnosis is not merely impractical but structurally incapable of converging on correct diagnoses in realistic clinical settings. Conclusion: Unless LLMs can independently collect patient history and triage, eliminate differential diagnoses, provide a treatment plan, and generate encounter notes, these models will not succeed in improving the efficiency of primary care delivery by human doctors. A different approach grounded in cognitive AI and structured reasoning is necessary. AI models should instead be seeded with weights provided by a panel of expert physicians to approximate an independent robot doctor.