In a qualitative study of 357 physicians published in the Journal of Medical Internet Research, physicians reported using unauthorized artificial intelligence tools to obtain second opinions on clinical decisions, support administrative work and professional development, and explore new technology.

The authors characterized this “Shadow AI” use as a response to gaps in formally approved, integrated, and accessible systems within health care organizations. "Shadow AI exists due to a combination of barriers, like rules and norms, and driving forces, such as curiosity, uncertainty, and interest in technology," first author Lena Petersson of Halmstad University in Halmstad, Sweden, told AACE Endocrine AI. "There is a need for more knowledge about its use in both clinical and administrative work."

The cross-sectional study surveyed physicians employed in Swedish health care organizations from December 2023 through January 2024. Of 557 eligible physicians invited from a verified online panel, 357 completed the survey. Eligibility required current use of artificial intelligence (AI) in clinical practice. 

Approximately 54% of respondents had 10 years or less of professional experience, 23% had 11 to 20 years, and 21% had more than 21 years. Most worked in the public sector: 54% were affiliated with public hospitals and 20% with public health centers. Fourteen percent worked in private health centers, 7% in academic or research institutions, and 3% in private hospitals. 

Researchers conducted conventional content analysis of physicians’ free-text responses on the use of unauthorized AI tools. The physicians’ descriptions predominantly involved general-purpose AI applications, particularly ChatGPT, accessed through personal accounts and private devices. None of the reported tools were certified as medical devices, procured through organizational channels, or integrated into clinical information technology infrastructure. The investigators therefore classified their use as Shadow AI based on the absence of both regulatory conformity assessment and organizational procurement, endorsement, and governance oversight. 

The analysis identified four categories of Shadow AI use: clinical work and decision-making, administrative work, research and professional development, and technological interest and curiosity. The study was not designed to estimate how frequently physicians engaged in each type of use. 

For clinical work, physicians described using Shadow AI as a "second opinion" or "colleague" for differential diagnosis, rare or complicated conditions, difficult-to-interpret symptoms, and diagnostic improvement. Reported needs also included risk prediction when intensive care is initiated, medication overviews, and decision support for treatment choices involving multiple variables. Physicians expressed interest in longitudinal patient analysis, although none reported having used AI for that purpose. One physician described entering deidentified data, such as medical history, examination findings, and test results, into ChatGPT to obtain possible differential diagnoses. Another reported occasionally using the tool to help identify rare conditions. 

Administrative uses centered on communication and documentation. Physicians reported using large language models to make specialized medical language or terminology more understandable to patients and relatives. One respondent described entering information from complex radiology reports into ChatGPT to produce more understandable patient letters. Physicians also identified potential uses for reminders involving medication, certificates, annual visits, and tests, as well as for optimizing visit scheduling and electronic health record documentation. 

In research and professional development, physicians used Shadow AI to obtain clinical knowledge, review current research, synthesize information, and support evidence interpretation. Respondents also raised concerns that excessive reliance on AI could cause physicians to forget their own clinical knowledge or prevent less experienced physicians from developing experience-based expertise. Limited time to learn how and when to use AI and a lack of local technical support were additional concerns. 

For the technological curiosity category, some physicians reported experimenting with AI to assess its capabilities and output, and several had programmed their own tools using ChatGPT or other open-source AI tools. Respondents also described an interest in evaluating AI before its formal implementation in health care. 

The authors emphasized that the four domains do not carry the same regulatory implications. Under the European Union Medical Device Regulation, software intended for diagnostic, predictive, monitoring, or therapeutic purposes requires conformity assessment before clinical use. The AI tools reported in the study lacked that assessment and were not formally procured or overseen by respondents’ organizations. 

The study had several limitations. The cross-sectional design captured AI use at a single point in time, and self-selection could have favored physicians with greater interest in technology. Because the analysis was qualitative, its purpose was to characterize Shadow AI rather than quantify the frequency of specific behaviors. The exclusively Swedish sample may also limit transferability to other countries. 

Free-text survey responses generally did not provide detailed information about prompting strategies, iterative exchanges with AI, verification practices, or integration into clinical workflows, which the authors said would require interviews or observational research. They also noted that rapid changes in AI could alter which tools are relevant to Shadow AI use over time. 

The authors noted that health care leaders should not view Shadow AI solely as a compliance violation but as a source of user-driven innovation signaling unmet professional needs. "The path forward involves creating secure, monitored environments where physicians can experiment with AI tools safely, bridging the gap between the speed of AI development and the pace of medical governance," wrote Petersson, and colleagues. 

The research was supported by the Business Models for Information-Driven Healthcare Ecosystems project, funded by the Swedish Knowledge Foundation, and was conducted as part of the Information Driven Care Research program at Halmstad University. Ingela Mauritzon is a doctoral student at the Multidisciplinary National Health Innovation Research School funded by the Knowledge Foundation. The funders had no role in study design, data collection, analysis, interpretation, or manuscript preparation. The authors declared no conflicts of interest.