Medical AI Raises Autonomy Questions Before It Reaches the Bedside
Medical AI is moving from an abstract promise to a visible part of clinical work, and the central question is no longer whether these tools will enter practice but how they will alter the physician’s role once they do. In a qualitative pretest published in JMIR Formative Research, investigators developed and tested a semistructured interview guide focused on the potential effects of medical AI systems on physician autonomy. The work was anchored in a seven-component model of autonomy proposed by Schulz and Harrison and used interviews with seven hospital physicians to examine how clinicians think about AI as a support tool, a constraint, or a possible threat to judgment.
The topic reaches beyond software performance. Autonomy shapes whether clinicians accept a technology, trust it in practice, and preserve ownership of the final medical decision. When AI enters diagnosis, triage, or treatment planning, it can alter workflow in subtle ways long before it ever appears to replace a physician outright.
Clinicians Accepted AI, But Not as a Substitute for Judgment
Most participants did not voice strong concern that AI would strip away their clinical autonomy. The dominant view was pragmatic rather than alarmist: AI could be useful if it remained an aid and did not become an authority. Physicians wanted systems that support reasoning, reduce cognitive load, and assist with routine analytic tasks without displacing the clinician’s responsibility for the final call.
One sentence cuts through the rest of the discussion: All physicians agreed that AI systems should not replace clinicians as primary clinical decision-makers.
That position matters because it places a clear boundary around acceptable use. In daily practice, a tool that recommends a course of action may be welcomed; a tool that quietly erodes the clinician’s independence is a different matter altogether. Physicians in the interviews recognized that AI can influence decisions even when it is labeled as advisory, especially if it becomes deeply embedded in routine documentation, triage, or ordering systems.
Deskilling and Automation Bias Emerged as the Main Risks
The autonomy concerns that surfaced were not limited to a dramatic loss of control. Several participants pointed to the more gradual danger of deskilling, where repeated reliance on algorithmic output could weaken a physician’s own diagnostic or interpretive capacity. In parallel, automation bias emerged as a meaningful risk: once a system produces a recommendation, clinicians may be more likely to accept it than to challenge it, especially under time pressure or in complex cases.
Those concerns are clinically familiar. A tool that appears objective can change the threshold for dissent, and a convenient recommendation can become a psychological anchor. In that setting, autonomy is not only about legal authority; it is also about preserving the mental space needed to question a machine when the bedside context does not fit the pattern the software expects.
Limited system explainability added another layer of unease. If clinicians cannot see why an algorithm reaches a specific output, then the physician is asked to endorse a recommendation without being able to test its internal logic against the patient’s presentation. That uncertainty can narrow meaningful autonomy even when the clinician retains formal authority.
Economic Pressure May Shape Autonomy as Much as the Algorithm Itself
The interviews also pointed to organizational forces that can compress autonomy from outside the exam room. Participants raised concerns about increasing economic or cost-related pressure, suggesting that AI could be used not only to improve care but also to standardize decisions in ways that prioritize efficiency, throughput, or resource control.
In practice, that creates a second-order autonomy problem. A physician may technically remain free to choose, yet the surrounding system may steer choices toward whatever is easiest to scale or cheapest to deliver. The pressure is subtle, but it can still change behavior. If an AI-supported pathway becomes the default, deviation may begin to feel exceptional even when it is clinically justified.
The study therefore framed autonomy as broader than a single decision point. Professional discretion, cognitive independence, and organizational freedom all entered the discussion. That wider framing is useful because the impact of AI is likely to appear first in workflow norms, not in formal declarations that clinicians have surrendered control.
The Interview Guide Captured Some Autonomy Dimensions Better Than Others
The second aim of the project was methodological: to assess whether the interview guide was adequate for exploring autonomy-related effects of medical AI. Investigators mapped the interview data onto the seven-component model of physician autonomy after first identifying themes inductively from participant responses. This hybrid approach allowed the team to test whether the guide captured the expected conceptual territory while also leaving room for unanticipated concerns.
The result was mixed in an instructive way. The interviews yielded meaningful material on autonomy threats, but the inductively identified themes and subthemes did not fully reflect all components of the underlying model. In other words, the guide was useful but incomplete. Some autonomy dimensions emerged strongly, while others were less visible in the responses generated by the current format.
That gap is not a methodological footnote. It suggests that future qualitative work on medical AI may need a more refined way to ask about autonomy so that the full construct is not reduced to a narrow conversation about control or replacement. A richer instrument will need to capture what physicians lose, what they retain, and what they quietly adapt when AI becomes part of routine care.
Why This Matters for Clinical Design and Future Research
The practical message is straightforward: medical AI is not being judged only on accuracy or efficiency. Physicians are also assessing whether it preserves their ability to think independently, challenge recommendations, and remain accountable for the patient in front of them. If systems obscure reasoning, push clinicians toward passive acceptance, or reinforce cost-driven decision pathways, autonomy may erode even in the absence of overt coercion.
The study offers an early framework for asking the right questions before broader implementation. It also shows that qualitative inquiry into AI should not stop at whether clinicians “trust” a tool. Trust is only one layer. Autonomy reaches into professional identity, cognitive habits, and the structure of clinical work itself.
A small sample of seven hospital physicians cannot settle those questions, and this pretest was not designed to measure real-world patient outcomes or quantify the frequency of autonomy loss. It did, however, reveal a clear direction for future research: the next generation of qualitative tools will need to probe how AI changes judgment, not just whether clinicians like it. The remaining challenge is to refine autonomy measurement before medical AI hardens into routine infrastructure and its influence becomes harder to separate from ordinary clinical practice.