Survey at Arizona finds 98% of clinicians use ungoverned AI tools for clinical tasks

The study, posted as a 2026 arXiv preprint, used a cross-sectional survey design to capture self-reported frequency across administrative, clinical, research, education, and teaching domains. Conversational AI and clinical decision support tools dominated usage, extending into higher-risk activities such as diagnostic assistance where outputs directly inform reasoning. This pattern reveals rapid adoption outside institutional governance, raising questions about data leakage and unvalidated model performance in patient-facing decisions.

Related work from the JAMA Network Open 2024 survey of 1,000 U.S. physicians and a 2025 NEJM AI analysis of ambient scribe tools shows similar off-platform use rates above 70 percent, often driven by workflow friction rather than deliberate risk acceptance. These studies collectively indicate that institutional policies lag behind individual tool discovery, creating inconsistent privacy protections and potential amplification of model biases when clinicians paste de-identified notes into consumer interfaces.

Next steps require mixed-methods studies that pair usage logs with output audits and patient outcome tracking. Without such data, health systems cannot design targeted training or approved alternatives that preserve efficiency gains while mitigating accuracy and confidentiality risks.

Evidence strength is limited by the small single-site sample and reliance on recall; multi-center longitudinal designs with objective telemetry would substantially strengthen causal inferences about downstream effects.