Most physicians believe artificial intelligence could improve their practice, yet fewer than one in three have actually used it in patient care. A cross-sectional survey analyzing responses from 1,049 physicians across 50 countries and territories found a significant disconnect between enthusiasm and real-world adoption. Only 27.8% of respondents had used AI clinically, and just 17.7% had received formal training in the technology. This gap reveals a critical challenge as healthcare systems worldwide race to integrate AI into clinical workflows.

What's Driving the AI Adoption Gap in Healthcare?

Healthcare is adopting AI at three times the rate of other industries, according to Raghav Mani, director of Digital Health at Nvidia. Despite this accelerated pace, the physician survey suggests that awareness and access to AI tools are not translating into widespread clinical use. The disconnect points to several barriers: lack of formal training, uncertainty about how to integrate AI into existing workflows, and questions about whether passive AI recommendations actually improve patient outcomes.

A randomized trial called PRIMA-AI tested this assumption directly in kidney transplant care. Researchers gave 76 kidney transplant recipients either usual care or access to an AI system integrated into their electronic health records (EHR) that predicted graft-loss risk. The finding was sobering: simply making AI risk estimates available to clinicians did not improve conversations about graft-loss options or shared decision-making outcomes. This suggests that deploying AI tools alone, without thoughtful workflow integration and clinician engagement, may not deliver the promised benefits.

How to Bridge the AI Training and Adoption Gap

• Formal Training Programs: Establish mandatory AI literacy and clinical application training for physicians before deploying AI tools in clinical settings, moving beyond the current 17.7% with formal training.
• Workflow Integration: Design AI systems that fit naturally into existing clinical routines rather than requiring clinicians to adopt entirely new processes or decision-making frameworks.
• Clinician Feedback Loops: Involve front-line users, clinical staff, and patients in AI implementation planning to ensure tools address real clinical needs and build trust in recommendations.
• Clear Evidence of Benefit: Conduct rigorous validation studies showing that AI recommendations actually improve patient outcomes, not just provide additional data points.

The Mount Sinai Health System's approach offers a model for moving forward. Brendan G. Carr, CEO of Mount Sinai Health System, described AI as a "new partner" to aid clinicians in synthesizing vast and growing clinical data. However, this partnership requires more than software deployment. Dave A. Chokshi, a physician and former New York City health commissioner, emphasized during a recent healthcare AI symposium that "it makes relationships even more important that we know then are". He stressed that building trust with patients and integrating feedback from front-line users will be vital as AI becomes more prevalent in healthcare.

"The real question is not IF AI will transform healthcare, but HOW," stated Girish N. Nadkarni, a nephrologist and practicing clinician at Icahn School of Medicine at Mount Sinai.

Girish N. Nadkarni, Nephrologist and Practicing Clinician, Icahn School of Medicine at Mount Sinai

The symposium also highlighted the importance of external validation and standardization. Andrew Gruen, standards lead at MLCommons, spoke about the need to establish benchmarks for AI use in research and healthcare settings. He emphasized that AI models require not just training but external evaluation and validation before clinical deployment. This rigorous approach contrasts sharply with the current landscape, where many AI tools enter clinical use with limited evidence of real-world benefit.

Another critical barrier is the complexity of clinical context. Researchers at BMJ Digital Health argued that clinical AI performance depends heavily on selecting and structuring the right patient information. This process, called "context engineering," involves retrieving, processing, and managing clinical data to reduce noise and outdated information. Without proper context management, AI systems may generate recommendations based on incomplete or irrelevant data, undermining clinician confidence.

The path forward requires a shift from viewing AI as a standalone tool to understanding it as part of a broader clinical ecosystem. Azra Bihorac, senior associate dean for research at the University of Florida, noted that while AI is continuously improving in its ability to assess problems and suggest next steps, human input remains vital for collaborative success. This human-centered approach acknowledges that clinicians bring irreplaceable judgment, intuition, and accountability to patient care.

As healthcare systems continue investing billions in AI infrastructure, the physician survey and clinical trial results suggest that training, workflow integration, and evidence of benefit must keep pace with technology deployment. Without these elements, AI adoption will remain limited to early adopters, leaving the majority of clinicians skeptical about whether the technology truly improves patient care.