The most revealing number about AI use in the medical field today is not that physicians are avoiding it. It is that they are using it while still signaling that the institutional floor underneath them is unfinished. In Doximity’s 2026 physician survey, 81% of physicians reported using AI, while 71% named accuracy and reliability of AI outputs as their greatest barrier or concern. Only 8% said they had clear institutional AI policies they understood, and 47% said their organization’s policies were still evolving.[1]
That is not a picture of simple resistance. It is a picture of normalization without settled governance. Separate organizational evidence points in the same direction: federal data on hospitals’ use, evaluation, and governance of predictive AI from 2023 to 2024 show that institutional adoption and oversight questions were already moving together before the current generative AI surge fully entered everyday clinical conversation.[2] In the broader 2026 adoption landscape, the headline figures often cited together—81% physician use and 70% organizational deployment—should be read carefully because they come from different survey populations and definitions of “AI use.” They still point to the same operational reality: AI is no longer waiting outside the health system.
Patients are not outside this contradiction either. Dialog Health’s 2025–2026 survey data found that 65.8% of US adults had low trust that health systems would use AI responsibly, and 57.7% had low trust that AI tools would not cause harm. Yet 52% of patients were already using AI for health research, while 78% expected doctors to validate that information.[3] The public is experimenting and asking for a safety layer at the same time.
Adoption Is Not the Same as Confidence
High use can look like endorsement when it is presented on an executive dashboard. In clinical operations, it often means something narrower: the tool is available, the workflow is pressured, and the clinician has found a way to make it useful enough while compensating for its uncertainty.
That distinction matters because healthcare AI is not one category. Ambient documentation, patient-message drafting, clinical summarization, predictive risk tools, diagnostic support, coding assistance, and public generative AI chatbots raise different questions. A physician who accepts an ambient scribe for note drafting may still distrust an AI-generated diagnostic suggestion. A nurse who uses an internal summarization tool may still reject a public chatbot for patient-specific clinical advice. Blending these uses into one adoption number makes the market look cleaner than the clinical reality.
The Doximity data are useful because they place use and concern side by side. Physicians are not waiting for a perfect policy environment before touching AI, but their leading concern remains the accuracy and reliability of outputs.[1] That concern is not abstract. It decides whether a clinician can trust a draft note, whether a risk flag deserves attention, whether a generated explanation can be passed to a patient, and whether a recommendation needs a second source before it enters care.
There is also a measurement problem. Physician satisfaction with AI is not the same as clinical outcome evidence. Doximity reports that 69% of physicians who use AI say it has improved care, but that may reflect the experience of physicians already inclined to try these tools rather than a generalizable finding about patient outcomes.[1] It is still meaningful; it just should not be asked to carry more weight than it can bear.
The Policy Gap Lands on Clinicians First
The 8% policy figure is the sharpest warning in the physician data. If only a small share of physicians say they have clear institutional AI policies they understand, then the system is asking clinicians to make judgment calls in a space where the rules are still half-visible.[1] That does not mean every use is unsafe. It means too much of the safety work is being relocated from institutional design to individual discretion.
| Signal | What it measures | Governance implication |
|---|---|---|
| 81% of physicians use AI | Reported physician adoption | AI is already inside routine work, regardless of whether policy is mature |
| 71% cite accuracy and reliability as the top concern | Physician concern about outputs | Use has not erased doubt about clinical dependability |
| 8% report clear institutional AI policies they understand | Perceived policy clarity | Clinicians lack a shared operating rulebook |
| 47% say policies are still evolving | Institutional policy maturity | Governance is catching up after deployment has begun |
In implementation work, unclear policy rarely stays theoretical. It shows up in small questions that determine whether an AI tool is governed or improvised: May a clinician paste a patient message into a public model? Can an AI-generated discharge instruction be sent after editing? Who approved the model embedded in a vendor platform? What documentation is required when AI influences a decision? Who reviews complaints when a patient challenges an AI-assisted communication?
When those answers are not visible, clinicians often create local workarounds. Some workarounds are harmless and practical. Others become shadow AI: unvetted tools used outside institutional oversight because the sanctioned path is unclear, slow, or absent. The risk is not only that a public tool might produce a poor answer. It is that no one knows it was used, no one has reviewed whether the use was appropriate, and no one can monitor the pattern before it becomes routine.
This is where education alone is too small a remedy. Training can tell staff what AI is and where it can fail. It cannot substitute for a policy that says which tools are approved, which data may be entered, which outputs require verification, which uses are prohibited, and who owns escalation when something goes wrong.
Validation Is Already Becoming the Informal Safety Layer
Clinicians are not passively accepting AI output. Wolters Kluwer’s 2026 Future Ready Healthcare survey found that 77% of clinicians said they always or often validate AI-generated health information. The same survey found that 92% of doctors and 90% of nurses said it was very or somewhat important for AI systems to be validated by a human expert in the loop.[4]
That behavior deserves more respect than it usually gets. It is not merely reluctance to modernize. It is clinicians building a safety layer around tools whose institutional rules, performance boundaries, and accountability chains may not yet be clear enough. The problem is that informal validation is uneven by nature. One physician may double-check every generated medication explanation. Another may review only outputs that feel surprising. A third may assume a tool embedded inside the electronic health record has already been fully vetted.
Human-in-the-loop language can also become decorative if it is not operationalized. A health system has not solved validation by saying a clinician remains responsible. It has to define what validation means for each use case. Reviewing an AI-drafted note for tone and completeness is not the same as validating a risk score that may influence monitoring intensity. Checking a patient-facing summary is not the same as adjudicating a diagnostic suggestion.
The useful governance question is not whether a human is somewhere near the loop. It is whether the right human is assigned the right kind of review at the right moment, with enough information to understand what the AI did and enough authority to override it.
Patients Are Asking for the Same Thing in Different Words
Patient distrust is often treated as an education problem: if people understood AI better, they would become more comfortable with it. The survey evidence points to a more demanding interpretation. Patients are already using AI for health research, and many still expect doctors to validate what they find.[3] That is not ignorance. It is a request for accountable interpretation.
The privacy signal is just as important. Dialog Health reports that 77% of patients are concerned about privacy in AI-driven healthcare.[3] A patient does not need to understand model architecture to ask a legitimate question: Where is my information going, and who is responsible if it is used in a way I did not expect?
Trust also depends on whether patients can tell when AI is involved. If a message is drafted by AI and approved by a clinician, the patient may care less about the drafting tool than about whether the clinician actually reviewed it. If a triage recommendation is influenced by an algorithm, the patient may want to know how to challenge it. Transparency does not require turning every encounter into a technical briefing. It does require enough disclosure for patients to understand when AI is materially shaping their care experience.
Why Governance Fits the Problem Better Than Education Alone
Education helps clinicians recognize hallucinations, automation bias, privacy exposure, and overreliance. Patient education can explain why an AI-drafted message still requires clinician review or why a risk tool supports rather than replaces judgment. But neither group can be educated into trusting a system whose rules are invisible.
Governance answers a different set of questions. It defines acceptable use before clinicians have to improvise. It assigns ownership for validation and monitoring. It separates low-risk administrative assistance from higher-risk clinical decision support. It creates a process for reviewing vendors and internally built tools. It gives compliance, clinical leadership, IT, and frontline staff a common reference point when the next tool appears.
A practical AI governance program does not need to freeze innovation while committees perfect every rule. It does need enough structure to prevent each department from inventing its own standard. At minimum, health systems need a live inventory of AI tools, a use-case classification process, data-use rules, validation expectations, monitoring plans, patient-communication standards, and an escalation route for safety, privacy, or equity concerns.
The federal predictive-AI governance context is useful here because it shows that the issue is older than the current fascination with generative models.[2] Hospitals were already facing questions about evaluation and oversight for predictive tools. Generative AI has made the governance gap more visible because the tools are easier to access, easier to use outside approved channels, and harder for institutions to observe when staff self-serve.
What Clearer Governance Changes in Daily Work
The goal is not to turn every clinician into an AI auditor. It is to remove avoidable ambiguity from clinical work. A physician should know whether an AI-generated patient summary can be copied into the chart. A nurse should know whether an AI translation or explanation tool is approved for patient communication. A department chair should know who is tracking model performance after deployment. A compliance officer should know which tools are processing protected health information and under what agreements.
- Approved-use rules: which AI tools may be used, by whom, and for which tasks.
- Data boundaries: what patient, employee, operational, or research data may be entered into each tool.
- Validation requirements: when a clinician must check AI output and what that review must include.
- Monitoring ownership: who tracks performance, incidents, drift, complaints, and equity concerns.
- Disclosure standards: when patients should be told AI contributed to a communication, recommendation, or workflow.
These are not abstract compliance artifacts. They change what happens at the point of care. They reduce the pressure on individual clinicians to guess. They make it easier to say no to unsafe shortcuts. They also make it easier to say yes to useful tools, because approval no longer feels like a quiet transfer of risk to the person holding the license.
Trust Will Vary by Use Case
It would be a mistake to describe healthcare AI as uniformly distrusted. Some uses are already routine, useful, and less controversial. Administrative drafting, documentation support, and summarization often feel different to clinicians and patients than diagnostic reasoning or autonomous triage. The governance burden should reflect that difference.
That is why broad AI policy statements are rarely enough. A health system needs risk-based distinctions that clinicians can actually apply. A low-risk documentation tool may require review for accuracy and completeness. A tool that influences diagnosis, treatment prioritization, or patient access needs a higher bar for validation, monitoring, and appeal. A public generative AI tool that has not been vetted for patient data should not be treated as a harmless convenience simply because it is easy to open in a browser.
Experience may improve trust over time. As clinicians use certain tools repeatedly and see where they help, skepticism can become more specific and less global. That is a healthy direction, provided familiarity does not replace evaluation. A tool can become comfortable before it becomes well governed.
The Trust Gap Is an Accountability Gap
The current paradox is not that physicians and patients reject AI while secretly using it. The paradox is that both groups are adapting faster than institutions are making accountability visible. Physicians validate outputs because they know the tool can be useful and wrong. Patients use AI for research and still want doctors to check it because they know information is not the same as care. Both behaviors point toward the same requirement: AI needs an accountable clinical and institutional frame around it.
Health systems will not rebuild trust by asking clinicians and patients to become more comfortable with AI in the abstract. Trust grows when policies are visible, use cases are defined, humans remain meaningfully in validation loops, performance is monitored after deployment, and unvetted AI stops being the easiest option.
References
- 2026 State of AI in Medicine Report, Doximity, 2026.
- Hospital Trends in Use, Evaluation, and Governance of Predictive AI, 2023-2024, HealthIT.gov.
- AI Healthcare Statistics, Dialog Health.
- 2026 Future Ready Healthcare, Wolters Kluwer, 2026.
Comments
Join the discussion with an anonymous comment.