Assessment of Non-Institutional AI Tool Usage Among Clinicians

Introduction

• AI use raises concerns about deskilling, security, and privacy, including risks from tools that may use inputs for model improvement or expose data.Prior studies reported reduced skill-assessment performance after delegated coding and reduced physician decision accuracy with AI use.Source: Introduction · pp. 1–2
• Few studies have described which AI tools clinicians use for work, the tasks they support, or how often they are used; non-institutional tool use had not been assessed.Non-institutional tools are not provided, managed, or governed by the institution and are typically not strictly regulated.Source: Introduction · p. 2
• Existing work addressed AI education, model selection, integration frameworks, and clinician competencies, but not clinicians’ current daily use.The authors characterize this as a disconnect between envisioned AI use and current practice.Source: Introduction · p. 2
• The study aimed to characterize clinicians’ non-institutional AI use, offer practical considerations for responsible use, and identify next research steps for clinician-AI interaction.Source: Introduction · p. 2

Data Collection and Ethics

• The University of Arizona and Banner Health institutional review boards approved the data collection and study methods.Source: Data Collection and Ethics · p. 2

Study Sample

• Eligible participants were medical students, residents, fellows, faculty, and advanced practice practitioners affiliated with COM-T or Banner, contacted once from May 20 to June 26, 2026.Incomplete surveys and fully completed surveys with all responses blank were excluded.Source: Study Sample · p. 2

Survey Details

• The survey collected background information, AI tools used, and tasks performed, including job title, department, career stage, and workload shares across five categories.Source: Survey Details · pp. 2–3
• Participants selected AI tool categories and task frequencies over the prior six months; Oracle Clinical AI Agent was the only institutionally available tool and was excluded from non-institutional analyses when appropriate.Task frequency used a seven-level scale from never to daily.Source: Survey Details · p. 3

Survey Analysis

• Because the sample was small, the analysis used descriptive statistics and comparisons rather than formal statistical testing.Categorical variables were summarized as counts and percentages, and continuous variables as medians and interquartile ranges.Source: Survey Analysis · p. 3

Characteristics of Study Sample

• The 44 respondents were mostly APPs or faculty with more than 10 years since training; Emergency Medicine and Medicine were the most represented departments.Median workload shares were 12.5% administrative, 65.0% clinical, 1.0% research, 1.0% studying/continued education, and 5.0% teaching; these medians defined the higher-versus-lower workload comparisons.Source: Characteristics of Study Sample · p. 3

Assessment of AI Tool Usage

• 28 respondents used each of Conversational AI and Clinical Decision Support/Diagnostic AI; 18 used each of OS-Integrated AI and Clinical Documentation and Workflow Automation tools.Source: Assessment of AI Tool Usage · p. 3
• APPs had a smaller proportion of AI users than medical students, residents/fellows, and faculty across surveyed tools, except that APPs had the highest proportion using Clinical Documentation tools.No clear patterns by career stage or department were observed.Source: Assessment of AI Tool Usage · p. 4

Assessment of Task-Based Usage

• 43 of 44 respondents (97.7%) used AI for at least one work-related task in the preceding six months, and administrative and clinical tasks showed the most frequent use, including daily use by multiple participants.The reported use remained after excluding the institutionally available tool and its associated tasks.Source: Assessment of Task-Based Usage · p. 4
• All respondents reported administrative and clinical AI use, including those spending below the median workload share, while research, studying/continued education, and teaching use occurred among those at or above the median.Source: Assessment of Task-Based Usage · p. 4
• Over 60% used AI for writing emails or messages and drafting letters, memos, or reports; clinical use most commonly involved drafting patient notes and transcribing visits.Clinical respondents at or above the median workload share also frequently used AI for diagnostic assistance and clinical decision support.Source: Assessment of Task-Based Usage · p. 4
• Research AI use was uncommon except for literature searching or summarizing, reported by 59%; studying use varied by task, and nearly all teaching use came from those above the median teaching share.One below-median teaching respondent used AI daily to simplify complex topics and generate handouts.Source: Assessment of Task-Based Usage · p. 4

• The survey characterized AI use across multiple clinical departments, career stages, and workload distributions.Source: Discussion · p. 4
• It assessed how frequently clinicians used AI tools for different work-related tasks.Source: Discussion · p. 4
• Tasks were assessed across administrative, clinical work, research, studying/continued education, and teaching categories.Source: Discussion · p. 4

• 43 of 44 respondents reported non-institutional AI use for at least one work-related task in the preceding six months.Source: Clinicians and Non-Institutional AI Use · p. 4
• AI use spanned all five workload categories and every task assessed, with administrative and clinical work used most frequently, especially for writing and documentation.Source: Clinicians and Non-Institutional AI Use · pp. 4–5
• Respondents also used AI for diagnostic assistance and clinical decision support, while few usage patterns varied descriptively by career stage, department, or workload.Source: Clinicians and Non-Institutional AI Use · p. 5

• For tasks even tangentially related to patient care, clinicians should independently validate AI outputs because tools can hallucinate responses.Source: Implications for Responsible AI Use · p. 5
• For lower-risk productivity tasks such as email writing, clinicians can take a more hands-off approach by reviewing outputs for errors.Source: Implications for Responsible AI Use · p. 5
• Institutions should define policies for non-institutional AI that address appropriate use, privacy, security, and accountability.Source: Implications for Responsible AI Use · p. 5
• For high-risk activities, clinicians should evaluate the full range of risks and benefits and eschew AI use if necessary.Source: Implications for Responsible AI Use · p. 5

• Research should establish what data clinicians enter into non-institutional AI tools and how they evaluate and verify generated outputs.Source: Future Directions for Clinician-AI… · p. 5
• Studies should assess short-term benefits or drawbacks, particularly for critical thinking, judgment, and skill degradation.Source: Future Directions for Clinician-AI… · p. 5
• Long-term research should examine how, why, and under what conditions AI can be safely and effectively integrated into clinician workflows.Source: Future Directions for Clinician-AI… · p. 5
• Education research should investigate how best to incorporate AI-specific training into medical education.Source: Future Directions for Clinician-AI… · p. 5

Limitations

• The response rate was about 2.2%, and possible self-selection plus a small, uneven sample prevented meaningful subgroup analyses.The authors note that clinicians with greater AI experience or interest may have been more likely to participate.Source: Limitations · p. 5
• Because the survey was conducted at one urban medical institution, its findings may not generalize to other healthcare systems or clinical settings.Source: Limitations · p. 5
• Usage was self-reported and not independently verified, while the brief survey omitted topics including tool-task relationships, trust, perceived quality, and personal use.The survey’s brevity was intended to support a higher response rate.Source: Limitations · pp. 5–6

Conclusion

• The descriptive survey examined non-institutional AI use among clinicians at a single academic medical center.Source: Conclusion · p. 6
• Respondents reported AI use across administrative, clinical, research, studying or continuing education, and teaching activities.Source: Conclusion · p. 6
• The authors identified practical considerations for responsible clinician AI use and future research on clinician-AI interaction in healthcare delivery.Source: Conclusion · p. 6

Administrative

• Administrative entries label letters or reports, repetitive-task automation, and scheduling as low risk, while policy drafting and meeting preparation are medium risk.Source: Category Task Potential Risk · p. 10

Clinical Work

• Clinical entries assign high risk to drafting patient notes, answering patient questions, diagnostic assistance, and clinical decision support.Source: Category Task Potential Risk · p. 10

Research

• Research entries label literature searching or summarization as low risk and manuscript editing, grant writing, and methodological review as medium risk.Source: Category Task Potential Risk · p. 10

Studying, Continued Education, and Teaching

• Educational and teaching examples span low-risk study aids and lecture summaries, plus medium-risk guideline comparisons and teaching materials.Source: Category Task Potential Risk · p. 11

Figure 1

• Figure 1 shows respondent counts by task and frequency, with bars split by whether workload allocation was at or above the category median.Frequency tags range from never to daily.Source: Figure 1 · p. 13