The integration of artificial intelligence into clinical research has sparked a critical debate regarding the future of medical evidence. In an opinion piece published in *PLOS Digital Health*, researchers Shenwei Wan of Renmin University of China and Xiaoguang Lei of the First Affiliated Hospital of Kunming Medical University explore whether AI will democratize the production of clinical evidence or merely solidify existing power structures.
The Challenge of the "Evidence Monopoly"
Evidence-based medicine was designed to balance clinical expertise with rigorous research. However, the authors note that the generation of authoritative evidence—particularly randomized controlled trials (RCTs)—is currently restricted by high costs and infrastructure requirements. Between 2015 and 2016, the median cost for a pivotal trial supporting a U.S. drug approval reached US$19 million.
This financial barrier has led to a geographic concentration of research. Data covering 784,585 trial sites between 2006 and 2012 revealed that 83 percent of these sites were located in 25 high-income OECD nations, while 91 low- and lower-middle-income countries hosted only 5 percent. Wan and Lei define this structural concentration as an "evidence monopoly," where the ability to generate data is limited to well-funded institutions, ultimately shaping global reimbursement policies and clinical guidelines.
AI as a Tool for Synthesis and Risk
Artificial intelligence offers potential efficiencies in evidence synthesis, such as automating literature searches, citation screening, and data extraction. A scoping review cited by the authors identified 388 AI tools across 137 studies. Despite this proliferation, the authors warn that these tools do not yet demonstrate a systemic reduction in research inequality or overall costs.
Furthermore, the use of AI introduces significant analytical hazards:
- Fabricated Claims: Large language models may generate summaries containing inaccuracies or hallucinations.
- Automation Bias: Researchers may over-rely on automated outputs, leading to search omissions or inconsistent data extraction.
- Linguistic and Geographic Bias: AI models often favor English-language publications and may exhibit narrow geographic applicability, as seen in depression detection tools that failed to account for diverse populations.
Toward an Evidence Commons
To mitigate these risks, Wan and Lei propose a shift toward an "evidence commons"—a framework where a broader range of institutions can inspect, challenge, and update clinical evidence under shared standards. They argue that this is particularly necessary for digital health, where software updates and algorithmic drift mean that a single RCT is insufficient to capture ongoing performance.
The authors suggest a "design fitness" approach, where methodological choices—such as implementation science or target-trial emulations—are matched to the specific research question. They recommend that academic journals mandate formal disclosure statements for AI-assisted research, requiring authors to document software versions, input data, validation protocols, and human verification steps.
Ultimately, the authors conclude that technical refinements alone cannot resolve the global geographic imbalance in clinical research. They advocate for international funding bodies and governments to invest directly in local clinical ecosystems—including laboratory infrastructure and regional regulatory systems—to ensure that AI serves as a tool for equitable participation rather than a mechanism for further centralizing the production of medical knowledge.

