Clinicians can now reach generative artificial intelligence (AI) tools at the point of care without enterprise procurement or electronic health record integration. Organizational surveys capture system-level adoption, but they may miss this clinician-initiated use. In this issue, Olszyk reports a voluntary, cross-sectional electronic survey of physicians, advanced practice providers (APPs), nurses, pharmacists, and other professionals at LifeBridge Health, Maryland. The survey was conducted from May 8 through August 20, 2026. Its aim was to find gaps between frontline AI use and the training, verification, patient communication, and governance that responsible adoption requires.
Of 182 initiated surveys, 152 were complete (83.5%). Respondents were mostly registered nurses (33.6%) and physicians (32.9%), followed by APPs (13.2%), pharmacists (10.5%), and other roles (9.9%). Overall, 83 respondents (54.6%) reported any clinical AI use. Only 27 (17.8%) had formal AI training, but 127 (83.6%) wanted it. Use differed by professional role (χ2₄ = 28.65; P < .001; Cramér V = 0.43), ranging from 80.0% of physicians to 29.4% of nurses.
Among the 83 users, 75.9% reported improved efficiency and 83.1% reported moderate or high trust in AI recommendations. Yet 79.5% had encountered incorrect output, and only 69.9% verified output often or always. AI influenced diagnosis, assessment, or management for 45.8% of users and medication-related decisions for 41.0%. Half (50.6%) selected external, nonenterprise applications. More than half of users (53.0%) never disclosed AI use to patients. Across all respondents, 61.2% favored limiting use to approved tools, 57.2% favored a hospital AI policy, and 75.7% rated governance important or an immediate priority. However, 61.8% were unsure whether AI belongs within clinical privileging, and only 12.5% favored a formal AI privilege with focused or ongoing professional practice evaluation.
The author proposes several safeguards that the survey did not test:
- A multidisciplinary review group that keeps a registry of approved tools, with intended and prohibited uses for each.
- Risk-tiered validation and monitoring of tools.
- Tool- and task-specific authorization rather than a blanket AI privilege.
- Training in verifying output and escalating errors.
- A practical definition of patient opt-out, with tool-specific consent for higher-risk uses.
The findings have several limitations. They come from a single health system and a voluntary sample that may overrepresent enthusiastic users. The invitation denominator is unknown, the data are self-reported, and some comparisons were exploratory. For physician leaders, the central message is that clinicians are already using AI. Governance, education, and patient communication need to catch up with that use rather than wait for enterprise deployment.
This summary is appearing ahead of print. The full article will be released in March/April 2027.
