The Missing Variable in AI Adoption
Health care's AI conversation has largely been a performance conversation — sensitivity, specificity, time saved, errors reduced. A new invited commentary in JAMA Network Open, responding to a qualitative study by Duong and colleagues, argues that this framing misses the variable that ultimately determines whether AI tools survive contact with real patients: trust. As the commentary states plainly, effective AI implementation "depends not only on technical performance, but also on public trust and acceptance." The concept at the center of this argument is "social license" — an informal, dynamic form of public acceptance that has to be earned and, crucially, can be withdrawn.
Three Conditions, Not One Checkbox
Drawing on scenario-based workshops with 34 health consumers in Queensland, Australia, Duong and colleagues identify social license as resting on three distinct dimensions rather than a single trust score. The first is relational engagement — whether the clinician-patient relationship, including shared decision-making and continuity of care, remains intact around the technology. The second is structural support — the governance scaffolding of data privacy, accountability, transparency, and equity. The third is performance reliability — whether the AI genuinely personalizes care, keeps clinicians in control, and functions accurately.
What makes this framework useful for clinical leaders is its context-sensitivity. In acute care, patients prioritize performance reliability, since speed and life-preserving accuracy matter most in the moment. In postacute settings, relational and structural concerns take precedence, because patients have the time — and the inclination — to scrutinize how a system is governed. The commentary frames this as a conditional stance: not a fixed yes, but "a 'yes, if… and when….'" This distinction helps resolve a puzzle in the existing literature, where some studies find performance drives AI acceptance and others find explainability and human oversight matter more — the answer, it turns out, depends on which clinical context is being studied.
The Pattern Holds at Scale
A single 34-person qualitative study could be dismissed as anecdotal, but the commentary is careful to triangulate it against much larger datasets. A survey of 3,000 U.S. adults found that trust in and choice of medical AI tracked closely with the presence of a clinician, formal governance or certification, and performance — a near-exact match to Duong et al.'s three dimensions. A separate multinational survey spanning 47 countries and 48,340 respondents found that public trust in health care AI was conditional on data privacy, accountability, accuracy, and human oversight. And a 43-country survey of 13,806 patients found a clear preference for explainable AI systems and physician-led decision-making — "even if it meant slightly compromised accuracy." Taken together, these findings suggest patients are willing to sacrifice some raw performance for transparency and a human still steering the decision.
This structural and performance framework also lines up with the European Commission's Trustworthy AI Principles — governance, safety, transparency, accountability, human oversight, and equity — which now underpin the EU AI Act's regulation of health care AI systems. The convergence across an Australian workshop, U.S. and multinational surveys, and European regulatory frameworks is, in the commentary's words, evidence that "the conditions surfaced by Duong and colleagues can be generalized beyond their sample."
A Working Example: AI Scribes
The commentary points to ambient AI scribes as a live case study in how social license operates in practice. Across six U.S. health systems, 30 days of ambient AI scribe use was associated with reduced self-reported clinician burnout, lower cognitive task load, less after-hours documentation, and more attention available for patients. This is precisely the kind of AI deployment that strengthens rather than displaces the clinician-patient relationship. But the same technology illustrates the fragility of that trust: because ambient documentation captures sensitive clinical encounters, the commentary notes that "trusted use depends on consent practices that are transparent, flexible, and genuinely voluntary." Performance gains alone do not guarantee acceptance if consent is treated as an afterthought.
Where the Evidence Runs Out
The commentary is candid about the limits of the underlying study. The Queensland workshops captured only English-speaking consumers in a single Australian state, responding to hypothetical scenarios rather than lived clinical encounters. The design also centered patient perspectives to the exclusion of clinicians, developers, regulators, and health-system leaders, and it asked participants to voice opinions in group settings — a structure that risks social desirability bias. Given that multinational data already show attitudes toward medical AI vary by demographic factors, health status, and prior technology use, the commentary suggests the specific conditions for social license likely vary by population and setting, and should be tested against real-world implementation rather than hypothetical vignettes alone.
What This Means for Practice Leaders
The operational takeaway for physicians, administrators, and health system leaders is that earning social license is not a one-time regulatory or marketing exercise — it has to be actively maintained. The commentary closes with concrete guidance: preserve the clinician-patient relationship as AI tools are introduced, invest in governance and performance safeguards calibrated to the specific clinical setting, and communicate those safeguards to patients in terms they can actually understand. For practices evaluating AI scribes, triage tools, or diagnostic support systems, the message is direct — a technically excellent tool deployed without transparency or a clinician in the loop is not guaranteed adoption. Trust, in this framework, is not a downstream consequence of good AI. It is a design requirement.
