Artificial intelligence could help more teams work with health evidence, but it cannot fix who gets studied or who has the resources to produce research. In an opinion article published Nov. 19 in PLOS Digital Health, Wan and Lei argue that a fairer evidence system depends on research capacity as well as technology.
Evidence-based medicine aims to combine the best available research with clinical expertise. Randomised controlled trials remain indispensable for many questions about benefits and harms, the authors write. But producing trials that influence health decisions takes funding, staff and infrastructure. The article cites an estimate that pivotal trials supporting US drug approvals in 2015–2016 had a median cost of US$19 million.
Research capacity is also geographically concentrated, according to figures the authors draw from a study of trial sites around the world from 2006 to 2012. Of 784,585 sites counted, 83% were in 25 high-income OECD countries, while 5% were in 91 lower-middle- or low-income countries. Where research is conducted matters because the patients and settings represented in trials can shape the evidence used for guidelines, procurement and reimbursement.
Wan and Lei distinguish an evidence hierarchy from an “evidence monopoly.” A hierarchy ranks study designs according to how well they answer a particular question and address bias. A monopoly, as they use the term, describes the concentration of the funding, data, expertise and infrastructure needed to produce authoritative evidence. They say it does not imply deliberate exclusion or a rejection of rigorous trials.
That distinction is relevant to digital health because an AI system can play different roles: it may be the intervention being tested, a tool for analysing routine-care data, or a way to synthesise research. A trial can evaluate one version of an adaptive system, the authors argue, but it cannot by itself establish how later software updates, workflow changes, alert fatigue or algorithmic drift affect performance. They call for external validation, implementation research and ongoing monitoring alongside trials.
The authors describe this as “design fitness”: choosing methods to fit the question. When randomisation is ethical and feasible, trials can be especially valuable for assessing efficacy and comparative effectiveness. Questions about uncommon harms, long-term safety and routine implementation may also need observational approaches. Quasi-experiments, implementation research, real-world monitoring and target-trial emulation can address other questions. These approaches are complementary, not interchangeable, and adding methods alone will not redistribute research resources.
AI could help with repetitive evidence work such as searching, screening, extracting, translating and updating information. The opinion cites a scoping review that identified 137 studies and 388 AI tools or platforms for evidence synthesis. That is evidence of experimentation, the authors caution, not proof that AI has cut costs or reduced inequality. Human checks and clear protocols remain important; AI does not replace methodological judgement or make a weak study design strong.
The tools also bring risks. The authors note that generated summaries can contain invented claims, searches can miss studies, and extraction can vary between runs. English-language dependence may leave out other settings, while model updates and automation bias can introduce further problems. And faster analysis cannot recover observations that were never collected: patients, clinics and languages absent from the evidence remain absent.
Wan and Lei argue that an AI-enabled evidence commons would let more institutions produce, inspect, challenge and update evidence under shared standards. They say that requires secure data access, independent evaluation, interoperable standards, trained local teams and community participation, alongside the capacity to collect data. Their article is an opinion and reports no original data collection or analysis; its numerical examples come from studies cited in the text.