Generative AI models like ChatGPT are transforming how doctors, nurses, and pharmacists learn clinical skills by creating dynamic virtual patients that adapt to each learner's questions in real time. A comprehensive systematic review of 15 studies involving 645 healthcare trainees found that AI-powered virtual patient simulations consistently outperform traditional scripted training methods, marking a significant shift in medical education.

The research, which synthesized evidence from five major medical databases through March 2026, examined how large language models (LLMs), a type of AI trained on vast amounts of text data, are being integrated into clinical education across nursing, medicine, pharmacy, radiography, and emergency medical training. The findings reveal both the transformative potential and the limitations of this emerging approach.

How Are GPT Models Being Used in Medical Training?

Virtual patients have long been a cornerstone of healthcare education, allowing trainees to practice clinical decision-making in safe, controlled environments without risk to real patients. Traditionally, these simulations relied on scripted, linear scenarios that followed predetermined pathways. GenAI, powered by large language models, fundamentally changes this dynamic by generating contextually relevant, adaptive responses in real time.

• Input Methods: Studies used text-based interactions (9 studies), voice-based conversations (5 studies), or hybrid approaches combining both modalities for more realistic patient encounters
• Output Formats: Virtual patients delivered responses as text (9 studies), spoken dialogue (5 studies), or both, with some incorporating 3D-embodied avatars (6 studies) while others used simpler text-based interfaces (9 studies)
• Model Selection: Thirteen of the 15 studies relied on OpenAI's GPT models, including ChatGPT, while one study used a fine-tuned model from another provider and one evaluated multiple model families including Claude and open-source alternatives

Unlike static simulations, these AI-driven systems can generate unique patient responses based on learner inquiries, simulate emotional responses, and adapt difficulty levels to match individual student progress. This represents a fundamental evolution from preprogrammed cases toward interactive, intelligent patient encounters that more closely mirror real clinical practice.

What Do the Study Results Actually Show?

The evidence supporting GenAI-powered virtual patients is compelling, particularly in controlled research settings. Among the 15 studies reviewed, six used rigorous experimental designs, including three randomized controlled trials (RCTs) where participants were randomly assigned to either AI-powered training or traditional methods.

In these controlled comparisons, GenAI-supported virtual patients consistently improved learning outcomes. One RCT with 21 participants found enhanced clinical decision-making compared to control groups. Another RCT involving 26 participants showed significant improvements in ophthalmology history-taking skills. A crossover RCT with 20 participants, where each student experienced both methods, demonstrated better medical history-taking performance with AI-powered simulations.

Beyond these specific trials, the broader evidence base examined user perceptions across 14 studies, communication skills in 4 studies, clinical reasoning in 3 studies, and overall performance in 7 studies. Learners consistently reported positive experiences with the technology, and performance metrics showed measurable skill improvements across multiple healthcare disciplines.

However, the researchers emphasized a critical caveat: the evidence base remains limited by brief intervention durations and predominantly single-session interactions. Most studies tracked learners over short periods rather than examining long-term retention or skill development over weeks or months.

What Are the Major Limitations Experts Identified?

Despite promising results, the systematic review identified several significant gaps that must be addressed before widespread implementation. The research highlighted that current GenAI-supported virtual patients struggle with emotional and behavioral complexity, meaning they may not fully replicate the nuanced, emotionally charged interactions that occur in real clinical settings.

Simulation adaptability remains another concern. While AI models can generate novel responses, they may not always adjust difficulty levels or learning pathways in ways that align with established educational theory. The review noted that the evidence base is characterized by a general lack of underpinning educational theory, meaning many implementations were not grounded in proven pedagogical frameworks.

Additionally, most studies lacked control groups or used validated measurement instruments inconsistently, making it difficult to isolate the specific benefits of AI-powered simulations from other factors. The heterogeneity in study designs, interventions, and outcome measures was so significant that researchers could not perform a meta-analysis, a statistical technique that combines results across multiple studies to identify broader patterns.

What Do Educators and Policymakers Need to Know?

The systematic review offers actionable guidance for institutions considering integration of GenAI-supported virtual patients into their curricula. The evidence supports feasibility and acceptability, meaning the technology is practical to implement and learners find it engaging. However, successful deployment requires addressing critical research and design gaps.

Educators and instructional designers should prioritize theoretical grounding, ensuring that AI-powered simulations align with established learning frameworks such as experiential learning theory or situated cognition. Institutions should also invest in longitudinal studies that track learner outcomes over extended periods, moving beyond single-session evaluations to understand how skills develop and persist over time.

Standardized design protocols are essential for ensuring consistency and quality across implementations. As more healthcare programs adopt this technology, establishing shared guidelines for how virtual patients should be designed, what interactions they should support, and how their effectiveness should be measured will accelerate progress and prevent costly missteps.

The reliance on OpenAI's GPT models in 13 of 15 studies also raises questions about vendor lock-in and the need for research exploring alternative model families, including open-source options that institutions could deploy independently.

Why Does This Matter Now?

Healthcare education faces persistent challenges: clinical hours are scarce, standardized patients are expensive and variable, and traditional simulations cannot adapt to individual learner needs. GenAI offers a scalable, consistent solution that addresses these constraints while potentially improving learning outcomes. The technology is already being adopted in real educational settings, making evidence-based guidance urgent.

As large language models continue to improve in sophistication and accessibility, the gap between what is technically possible and what is pedagogically sound will only widen. This systematic review provides a foundation for closing that gap, offering educators, researchers, and policymakers a roadmap for safe, effective implementation while identifying the critical research questions that must be answered in the coming years.