A Comparative Study of Generative Artificial Intelligence Versus Clinical Experts for Evidence-Based Decision-Making in Austere Environments

INTRODUCTION: The exponential growth of biomedical data hinders the rapid adoption of evidence-based practices in the Military Health System. This is critical in austere environments where teams rely on limited resources. Generative artificial intelligence (AI) offers a potential solution for rapidly synthesizing evidence. This comparative study evaluated the utility and feasibility of generative AI tools compared to human clinical experts in identifying protocols for surgical instrument reprocessing in austere settings. MATERIALS AND METHODS: We conducted a descriptive comparative study to query four AI platforms (NIPRGPT, ChatGPT, Google Gemini, GenAI.mil) and two clinical experts. The authors prompted each group to identify a single best recommendation for reprocessing surgical instruments in austere environments without steam sterilization capabilities. We compared the outputs based on time-to-completion, accessibility behind Department of War (DoW) firewalls, and clinical validity against a literature review. RESULTS: AI platforms generated recommendations in under 10 minutes. Clinical experts required 14 hours to review and synthesize data. Regarding accessibility, commercial platforms (ChatGPT, Gemini) were blocked by DoD firewalls, while GenAI.mil was accessible. Clinical experts recommended chlorine dioxide (ClO2) due to its sporicidal properties, which ensure sterility assurance. Only ChatGPT matched this recommendation. Conversely, GenAI.mil and Gemini recommended ortho-phthalaldehyde (OPA), and NIPRGPT recommended glutaraldehyde. The AI models prioritized processing speed over sterility assurance. CONCLUSIONS: Generative AI significantly reduces the cognitive load and time required to synthesize clinical protocols. However, government-hosted AI tools prioritized logistical factors over safety standards in this study. We identified an accessibility-accuracy paradox where the most accessible tool provided less rigorous safety recommendations. Implementation requires human verification and specific governance to ensure AI supports rather than replaces clinical judgment.