Impact of LLM assistance on physician decision-making: a multi-country randomized controlled trial

Abstract

Abstract Disparities in the quality of healthcare persist globally, with poor-quality care contributing significantly to preventable mortality, particularly in low- and middle-income countries. While digital technologies, including generative artificial intelligence (AI), hold promise for improving clinical decision-making, their global effectiveness and potential to mitigate cross-country variation remain underexplored. We conducted a parallel-group randomized controlled trial across three economically diverse countries—Indonesia, Kenya, and the Netherlands—to evaluate the impact of large language model (LLM) access on physician performance using standardized clinical vignettes. Physicians (N=249) were randomly assigned to either a control group or an intervention group with access to GPT-4o. Results showed that LLM access significantly improved clinical performance, with the largest effect in Kenya (18%, 95% CI: 12.7 to 23.2, p < 0.001), followed by Indonesia (10.7%, 95% CI: 5.7 to 15.7, p < 0.001) and the Netherlands (7.2%, 95% CI: 3.7 to 10.7, p < 0.001). Notably, LLM access reduced cross-country performance disparities, particularly between Kenya and the Netherlands. However, distributional effects varied, with increased score dispersion in Indonesia and reduced variation in Kenya. Higher LLM usage was associated with greater performance gains, though some physicians without access outperformed those with access, suggesting that effective use depends on individual engagement. Our findings demonstrate that LLMs can enhance clinical performance across diverse settings while potentially narrowing global inequalities in care quality. Further research should explore mechanisms of effective LLM integration and long-term impacts on real-world clinical practice.