The key finding
A 2026 review published in Expert Review of Pharmacoeconomics & Outcomes Research identifies four major interconnected challenges preventing AI-generated synthetic health data from being reliably used in healthcare decision-making. These challenges—bias propagation, the privacy-utility trade-off, lack of standardized human oversight, and underdeveloped regulatory frameworks—currently limit synthetic data’s use as standalone evidence for informing drug pricing, reimbursement decisions, and health policy. The authors conclude that synthetic data should be integrated with real-world evidence in a hybrid ecosystem rather than used alone, supported by equity impact assessments and coordinated regulatory guidance.
What the study looked like
This was a non-systematic narrative review examining how synthetic data is generated and applied in health economics and outcomes research (HEOR). The researchers conducted targeted searches of PubMed and Google Scholar, prioritizing publications from 2019 onward to capture recent developments in AI-generated health data. Rather than analyzing a specific patient population or conducting new experiments, the authors synthesized existing literature to create a conceptual framework. They focused specifically on understanding the methodological and practical challenges that emerge when artificial intelligence creates simulated patient data intended to substitute for or supplement real patient records in research that guides healthcare spending and policy decisions.
Why researchers think this happened
The four challenges identified are interconnected rather than isolated problems. Bias serves as the foundational upstream driver—if the real-world data used to train AI models contains biases (such as underrepresentation of certain demographic groups), the synthetic data will replicate and potentially amplify these distortions. The privacy-utility trade-off emerges because techniques that better protect patient privacy often reduce how useful the synthetic data is for research purposes. The absence of standardized human-in-the-loop evaluation means there’s no consistent way for experts to verify that synthetic datasets accurately reflect clinical reality before they’re used in consequential decisions. Finally, underdeveloped regulatory frameworks leave researchers and policymakers without clear guidelines on when and how synthetic data can appropriately inform decisions about which treatments get funded or how much they should cost. These gaps exist partly because synthetic data technology has advanced faster than the evaluation methods and governance structures needed to ensure its responsible use in high-stakes healthcare contexts.
How to read this carefully
This is a narrative review rather than a systematic analysis with pre-specified inclusion criteria, meaning the authors exercised judgment in selecting which studies to include. The paper doesn’t present new data but rather synthesizes existing perspectives and challenges. Importantly, the concerns raised reflect the current state of the field in 2026—ongoing research may address some of these limitations. The authors focus specifically on health economics and outcomes research, where synthetic data would influence resource allocation decisions; findings may not apply equally to other research contexts like basic science or clinical trials. Readers should note that the paper advocates for a hybrid approach rather than rejecting synthetic data entirely, suggesting the technology has potential if properly governed. The review doesn’t quantify how often current synthetic data methods fail or succeed in real-world applications.
What this means for everyday life
Given these findings, it’s worth being aware that AI-generated patient data may increasingly influence which medications your insurance covers and how much they cost—but this technology isn’t ready to drive such decisions alone. If you participate in health research or see studies based on synthetic data, consider asking whether the findings were validated against real patient outcomes and whether the original data included people with demographics similar to yours. For healthcare advocates concerned about equity, this review suggests pushing for transparency about how synthetic data is created and whether it adequately represents marginalized populations who are often underrepresented in medical datasets. The hybrid approach recommended here—combining synthetic data with real-world evidence—suggests that the most trustworthy health research in coming years will likely draw on multiple data sources rather than relying on AI-generated simulations alone.