AI-mediated relational competence is proposed here as the emerging clinician’s ability to preserve empathy and patient trust, improve care transparency and shared decision-making, and exercise equity awareness and accountable human judgment when AI tools have the potential to influence all the elements of the care encounter and the resulting treatment recommendations. The construct integrates existing scholarship on empathy education [19-22], trust in health care AI [14,15,18,26-31], AI competencies for clinicians [6-8,14,33], clinical communication, and competency-based medical education [34-37].

Three features distinguish this construct from existing AI literacy frameworks. First, it is relational rather than technical. Where AI literacy frameworks emphasize what the clinician must know about the algorithm, relational competence emphasizes what the clinician must do with that information on behalf of the patient [13,14]. Second, it is assessable rather than aspirational. The construct is operationalized through observable learner behaviors that can be taught, practiced, evaluated, and remediated within established competency-based assessments including objective structured clinical examinations (OSCEs), standardized patient encounters, direct observation, and/or portfolio assessment [35,37]. Third, it is integrative rather than additive. The construct is not proposed as a new soft skill bolted onto existing curricula. It is a clinical performance domain for technology-enabled practice that maps onto the 8 competency domains synthesized in the Physician Competency Reference Set, in particular, professionalism, interpersonal and communication skills, and patient care [36] domains, and it aligns with the AAMC competency framework currently in development [12].

A recent framework proposed a three-lever approach to concurrent reform of medical AI education, emphasizing that curricular, assessment, and regulatory mechanisms must move together [33]. The construct proposed here is intended to provide the conceptual content that such reform efforts can teach and assess.

The proposed 6 subcompetencies are related but not redundant, and the relationships among them are themselves part of the construct. They can be read along 2 dimensions (Figure 1). The first concerns the relational stance the clinician sustains with the patient, expressed through empathic presence, demonstrated trustworthiness, and equity-aware trust-building. The second concerns the clinician’s conduct toward the AI and its outputs within the encounter, expressed through transparency in AI use, relational judgment, and accountable human decision-making. Relational judgment functions as the hinge between the 2: the capacity to read when the tool is supporting or eroding the relationship and to act accordingly. Because the construct describes relational work, some conceptual overlap among the subcompetencies is expected rather than a defect; the encounter is a single performance, and these are facets of it rather than independent modules.

What justifies treating these facets as distinct assessable targets is that they dissociate in practice. A learner may disclose that AI informed the recommendation (transparency) while uncritically deferring to its output (a failure of accountable decision-making). A learner may communicate warmly and attentively (empathic presence) while overstating the authority of an algorithm (a failure of trustworthiness). Or a learner may explain a tool clearly to one patient while dismissing the well-founded skepticism of another (a failure of equity-aware trust-building). Because these behaviors can be present or absent independently of one another, each can be separately observed, taught, and remediated in a way that a single global rating of “relational skill” would obscure. The construct is therefore neither a relabeling of empathy nor a relabeling of AI ethics. Rather, it specifies the distinct and separately assessable behaviors through which a clinician keeps an AI-mediated encounter empathic, transparent, trustworthy, accountable, and equitable.

The relational dimensions of clinical care were well described long before AI entered the encounter. Competency frameworks have consistently named interpersonal and communication skills, professionalism, and patient care as distinct domains of physician performance [36]. Empathy education has established that the capacities underlying those domains are teachable and assessable [19-21], and the competency-based movement has supplied the assessment architecture through which they are evaluated [34,35]. What that literature has not required is a single organizing term. “Relational competence” is used here as an integrative label for capacities that existing frameworks distribute across several domains. It is proposed in this viewpoint rather than adopted from established usage, and the claim advanced is not that these capacities are new but that AI mediation changes the conditions under which they must be enacted and assessed.

The proposed construct is, therefore, best understood as a performance layer rather than a replacement for adjacent domains, each of which answers a different question. AI literacy concerns what the clinician knows about the tool: its inputs, limits, and failure modes. AI ethics concerns the normative questions the tool raises, including fairness, privacy, bias, and the just distribution of benefit and risk. Empathy education concerns the clinician’s recognition of and response to the patient’s experience. Professionalism concerns role fidelity and the standards of conduct to which the clinician is accountable. Calls for a data-informed health workforce sit alongside these as a further technical competency [38]. AI-mediated relational competence names none of these. It names the observable clinical behaviors through which all of them are enacted when AI participates in a specific encounter with a specific patient. Some conceptual overlap with its parent domains is expected, since the encounter is a single performance rather than a set of independent modules.

This raises a reasonable question about whether earlier clinical technologies required the same construct. Electronic health records, rule-based decision support, and telemedicine each altered the encounter, shifting clinician attention toward the screen and generating its own digital-health competency frameworks [39]. Some of these systems incorporated earlier forms of AI well before the recent widespread adoption of generative models. The distinction proposed here is therefore better understood as a matter of degree than of kind, turning on the extent of a system’s participation in the encounter rather than its mere presence. Earlier systems, AI-enabled or not, were principal systems in which the clinician queried a record or a rule set and remained the author of the note and thereby the source of the words the patient heard. Contemporary generative systems are more integrative and proactive. They generate the clinical record rather than store it [40]; they deliver recommendations already formed and carrying implied authority [14,27]; and they can produce patient-facing language that reads as empathic. Each shift relocates something that previously belonged to the clinician. It is that relocation, rather than screen time or workflow disruption, that the construct is designed to address. The technologies principally at issue here are therefore ambient documentation tools, generative language models, and predictive or diagnostic decision-support systems, rather than digital tools generally.

Competency-based medical education requires observable behaviors rather than aspirational values [34,35]. Accordingly, this proposed construct is divided into 6 subcompetencies, each defined in terms of learner behaviors that can be assessed in clinical or simulated settings and evaluated against explicit performance criteria. Table 1 provides the full construct justification matrix. The subcompetencies are empathic presence, trustworthiness, transparency in AI use, relational judgment, accountable human decision-making, and equity-aware trust-building. Curricular maturity is treated separately as an institution-level dimension.

Empathic presence is the capacity to recognize, communicate understanding of, and respond to a patient’s experience while concurrently using digital or AI tools. It extends established empathy education [19-21] into AI-mediated contexts in which attention and conversational rhythm are altered by the presence of ambient scribes and other screen-mediated workflows [13,16,23]. Validated empathy measurement instruments, such as the Jefferson Scale of Physician Empathy [41], provide the psychometric foundations for assessing empathic orientation in medical learners but were developed prior to widespread AI integration and rely on self-report rather than observed behavior, thus offering a necessary but insufficient starting point for assessing empathic presence in AI-mediated encounters.

A further distinction is necessary as generative systems become capable of producing patient-facing language. Large language models can generate text that reads as warm, attentive, and responsive. Controlled work in medical education has shown that structured practice with role-prompted conversational agents can improve learners’ perceived communication competence across emotionally demanding scenarios [43]. Empathic presence, however, is not reducible to the production of sympathetic phrasing. It requires attention to what is occurring in front of the clinician. This requires sensitivity to context that the model cannot observe and to the professional accountability that attaches to a person rather than a system. AI-generated language may support communication and, in some settings, may improve it, but the clinician remains responsible for interpreting the patient’s situation, judging what the moment requires, and delivering care within a relationship for which someone can be held accountable. Assessment of empathic presence in AI-mediated encounters should therefore target the learner’s attentiveness and judgment rather than the fluency of the language produced.

Trustworthiness is the demonstrated capacity to act and communicate in ways that justify patient confidence. It draws on the expert consensus work, establishing trust as a multidimensional construct distinct from empathy [15], and on patient research, demonstrating that transparency, accountability, and respect for patient agency are central to patient trust in AI-enabled care [17,18,25].

Transparency in AI use is the ability to explain when and how AI informs care in patient-understandable language. This subcompetency reflects both the patient’s need for clear disclosure [17,25] and the clinician’s challenge of explaining AI-supported recommendations where the underlying logic may be opaque [26,27]. Conceptual work on AI-supported shared decision-making argues that effective shared decision-making in AI-enabled care requires clinically meaningful reasoning beyond just technical transparency [44].

Relational judgment is the ability to determine when AI use supports or undermines patient-centered care. AI tools have been shown to create distance between clinicians and patients, encourage an overreliance on AI outputs, and potentially amplify inequity [13,16,42]. Clinical learners must develop the skills to recognize these effects when they are present and then adapt the encounter accordingly.

Accountable human decision-making is the ability to maintain professional responsibility for recommendations shaped by AI. The literature on AI in clinical decision support has frequently identified the risk of responsibility diffusion when algorithmic recommendations enter the workflow [14,26,30]. Competency-based medical education must preserve clinician accountability, requiring learners to verify AI outputs and then convey that AI informs but does not replace clinical judgment. They must then document their reasoning in a way that maintains the clinician as the accountable decision-maker [35,36].

Equity-aware trust-building is the ability to recognize how bias, inequity, prior harm, or historical mistrust may shape patient responses to AI-mediated care. Bias and fairness concerns are central to AI ethics [2,14,42], and these concerns further shape relational legitimacy. Patients from communities with prior experience of medical bias may be more skeptical of algorithmic recommendations and of being distanced from their clinicians, and learners must be prepared to acknowledge and engage with that skepticism rather than dismiss it. The previously cited JAMA survey also documented that trust in health systems’ use of AI varies systematically across demographic groups, thus reinforcing the centrality of equity to relational trust-building [18].

Because trustworthiness, transparency in AI use, and accountable human decision-making all concern the clinician’s conduct in relation to the tool, their assessment boundaries warrant statement. Transparency concerns disclosure and whether the patient is told that AI informed their care in terms they can understand and use. Accountable human decision-making concerns verification and ownership and whether the clinician checks the output and retains responsibility for the resulting recommendation. Trustworthiness is broader than either and is assessed at the level of the encounter as a whole, namely whether the clinician’s conduct justifies the patient’s confidence, including respect for patient agency, consistency of care across patients, reliability over time, and candor about the limits of what the tool and the clinician can know. The distinction is observable in practice. A learner may disclose AI use fully and still not be trustworthy, for instance, by disclosing in terms that discourage questions, and a learner may verify an output diligently while conveying a false impression of certainty. Table 2 sets out illustrative competent and concerning behaviors for each subcompetency to support assessment at these boundaries. As with the maturity model and assessment blueprint described here, these behavioral anchors are proposed rather than empirically validated and require confirmation before use in high-stakes assessment or institutional benchmarking.

Three clinical vignettes illustrating the construct in practice, offered as proposed rather than empirically validated teaching examples, are provided in Multimedia Appendix 1: an ambient documentation tool that is running when a patient hesitates to disclose that cost has prevented them from taking a medication as prescribed; a patient who asks whether an AI-generated low-risk assessment reflects the tool’s judgment or the clinician’s; and a patient who asks whether a risk model was developed using data from patients like them. Each vignette identifies the subcompetencies it engages, contrasts a competent response with a plausible failure, and indicates where the scenario could be assessed.

Each subcompetency defined here can further be mapped to established teaching and assessment modalities [24,35,37]. Recent work has demonstrated the feasibility of AI-simulated patient interactions for training and assessing clinical interview skills [24], suggesting that AI-supported assessment infrastructure is actively developing in parallel with the educational challenges that AI itself creates.

The key empirical question raised by this construct is not whether medical schools mention trust, empathy, or even AI in their curricular outlines. Nearly all do in some form, whether in mission language, course and syllabus catalogs, or in stand-alone AI literacy sessions [7]. The more consequential question is whether schools systematically teach, reinforce, and assess these competencies.

This distinction is well-recognized in competency-based medical education scholarship. The transition from time-based to outcomes-based training has highlighted persistent gaps between what curricula declare and what learners can demonstrably do [34,35,37]. A recent study on the design and implementation of a digital competence framework for health professionals emphasized that competence assessment must be tied to defined frameworks and tested in real professional contexts rather than inferred from curricular inclusion [39]. Table 3 presents a 5-level curriculum maturity model for AI-mediated relational competence, ranging from implicit or absent (level 1) through introduced, integrated, and assessed to competency-based (level 5), with the latter defined by linkage to feedback, remediation, progression, and entrustment decisions.

The maturity model is proposed rather than validated. Its levels are derived from established distinctions in competency-based medical education between curricular mention, longitudinal reinforcement, formal assessment, and entrustment [34,35], but they have not yet been subjected to consensus or psychometric evaluation, and they should not be used for high-stakes curriculum evaluation or institutional comparison on the strength of this viewpoint alone. A validation program would proceed in stages: expert consensus review to confirm that the levels are exhaustive, mutually exclusive, and correctly ordered; cognitive interviewing with curriculum leaders to establish that the level descriptors are interpreted consistently across institutions; pilot coding of publicly available curriculum documents with interrater reliability testing to determine whether independent coders assign the same level to the same program; and systematic collection of validity evidence linking assigned levels to observable learner performance [37]. Until that evidence exists, the model is best used formatively as a shared vocabulary for institutional self-assessment and curricular planning rather than as a benchmarking instrument.

This framework is intended to serve 2 empirical purposes. First, it can support online curriculum assessment, in which publicly available curriculum materials are coded against the maturity levels to produce a national baseline picture of where US medical schools currently stand. Second, it can inform more systematic national survey research of medical and clinical programs. Comparing what schools publicly document with what curriculum leaders report is itself a useful empirical question.

A construct framework is a methodological precondition for both online curriculum assessment and national survey research. Without an explicit construct, coders cannot reliably distinguish symbolic from systematic inclusion, and survey items cannot be mapped to a coherent underlying competency model. The framework proposed here is designed to support both.

For online curriculum assessment, the construct provides a coding rubric that can be applied to publicly available materials and to the AAMC Curriculum Inventory. Each subcompetency from Table 1 generates a set of indicators (keywords, learning objectives, assessment modalities, and course types) that coders can search for and classify against the maturity levels of Table 3. This methodology mirrors published approaches to curricular content analysis in medical education.

For national survey development, the construct provides item domains corresponding to each subcompetency and maturity level. Items can ask not only whether a topic is taught but also whether it is reinforced longitudinally, assessed with explicit criteria, reflected upon for remediation, and linked to progression decisions.

Table 4 sets out an illustrative assessment blueprint, indicating how the subcompetencies might be distributed across settings and methods already available in most programs, from preclinical standardized patient encounters through workplace-based observation, simulation, and portfolio reflection. The blueprint is a worked example rather than a validated instrument. It has not been piloted; the mapping of subcompetencies to settings is one defensible arrangement among several; and the validation requirements stated for the maturity model apply to it equally. Its purpose is to show that the construct can be assessed using infrastructure that programs already operate, rather than to prescribe how any particular program should do so.

The competent and concerning behaviors in Table 2 correspond to the 6 subcompetencies defined above and are intended to complement, rather than duplicate, the maturity levels in Table 3. Table 3 describes how deeply a curriculum addresses the construct, while Table 2 describes what competent and concerning performance looks like within any given encounter.

Table 4 builds on both preceding tables by mapping the same subcompetencies onto concrete assessment settings and methods already available in most training programs.

Curricular maturity describes what an institution teaches. It does not describe what learners observe. The values and behaviors transmitted through routine clinical practice, rather than through formal instruction, are recognized influences on professional formation and plausible contributors to the erosion of empathy documented during training [22]. AI-mediated relational competence is exposed to the same dynamic. A learner taught to disclose AI use, verify AI outputs, and address patients’ concerns about algorithmic fairness will discount that teaching if supervising clinicians do not follow suit and reinforce these behaviors.

Faculty development is therefore a condition of the construct’s viability rather than an implementation detail. Supervising clinicians require preparation to model the behaviors specified in Table 1 by disclosing AI use in patient-understandable terms, seeking patient agreement where documentation tools are in use, verifying AI outputs before acting on them, recognizing automation bias in their own reasoning, and engaging concerns about fairness and reliability with the seriousness afforded other clinical questions. Evidence from AI-generated educational documentation is instructive here. Mischaracterizations and hallucinations occurred at low but nontrivial rates and were largely corrected when faculty reviewed the output before submission [40]. The findings illustrate the potential value and feasibility of faculty verification in that setting. An institution could rate highly on formal curricular provision while its clinical environment models the opposite, and self-assessment against Table 3 should be interpreted with that limitation in view.

AI does not render trust and empathy obsolete in medical education; rather, it raises the importance of both as foundational human inputs to the care encounter. As AI tools enter routine clinical workflows, the relational skills clinicians perform with patients become more challenging and also more critical to assess. The existing AI competency literature has identified the technical, evaluative, and ethical skills clinicians need [6-8,14,33], and the existing empathy education literature has similarly established that relational skills can be taught and assessed [19-22,41]. What has been missing is a competency framework that bridges these 2 bodies of work and operationalizes the relational competencies required for AI-mediated practice.

The construct proposed here and the accompanying curriculum maturity model are offered as the bridge. The framework shifts the question from whether trust and empathy are mentioned in medical curricula to whether they are taught, practiced, assessed, and required as observable performance in AI-enabled care. The rationale for this construct is intended to provide a conceptual foundation for subsequent empirical work.