AI as Teammate: Rethinking Task Distribution in Medical Training

Synthesis: Tsim et al. (2026) propose a conceptual reframing of AI integration in Medical and Health Professions Education: the problem is not learner misuse of generative AI but misclassification — a mechanistic failure of real-time metacognitive evaluation in selecting a subzone-inappropriate AI interaction mode. They introduce the SCAN framework (Substitute, Complement, Aid, Non-Negotiable), a human-centric decision scheme for generative AI task allocation grounded in Vygotsky's Zone of Proximal Development and Metacognition. The account yields testable predictions about how misclassification can be detected, mitigated, and prevented in the clinical learning environment, and it reframes skill acquisition (upskilling) and failure (de-skilling, never-skilling, mis-skilling) at the individual task level rather than via fixed-phase, cohort-wide treatments.

From Misuse to Misclassification

Integrating artificial intelligence — particularly Generative AI — into medical training has prompted widespread concern about learner over-reliance, misuse, and the erosion of foundational clinical competencies. Survey evidence suggests a large gap between exposure and readiness: a recent study found that while 97.1% of graduate medical trainees report no formal AI instruction, 85.5% already use AI tools for clinical decision support and academic writing.

The authors argue the dominant "misuse" framing misdiagnoses the problem. Instead, they propose that failures arise from misclassification — a failure of real-time metacognitive evaluation in which a learner selects a subzone-inappropriate AI interaction mode for a given task. The corrective is not restriction but more precise task classification.

The SCAN Framework

SCAN names four modes of generative AI task allocation grounded in Vygotsky's Zone of Proximal Development:

1. Substitute — AI performs a task the learner could perform independently; appropriate only when the capacity is already consolidated and substitution serves efficiency rather than bypassing formation.
2. Complement — AI works alongside the learner, providing support that extends but does not replace the learner's active contribution.
3. Aid — AI offers Scaffolding within the learner's zone of proximal development, supporting a capacity that is currently forming.
4. Non-Negotiable — tasks that must be performed by the learner without AI, where delegation would short-circuit the formation of foundational clinical competence.

The framework links each mode to the social-constructivist logic of the ZPD: the appropriate mode depends on where the individual learner stands relative to the task, not on a fixed phase or cohort-wide rule.

Skill Acquisition and the Triad of Skill Failure

The paper illustrates how trajectories of skill acquisition and failure operate at the individual task level:

1. Upskilling — the learner progressively internalizes the capacity, allowing AI support to be withdrawn or reclassified over time.
2. De-skilling — a previously formed capacity atrophies as AI takes over work the learner once performed.
3. Never-skilling — a capacity that was never formed because AI did the work from the start (the learner "cannot atrophy a muscle that was never built").
4. Mis-skilling — the learner forms a flawed or shallow capacity through inappropriate AI mediation.

The authors identify passive engagement within correctly classified AI-scaffolded tasks as a particularly insidious, detection-resistant pathway to mis-skilling. This requires re-identifying the subzone — shifting from AI assistance to expert assistance, with human experts serving as epistemic auditors.

What the reframing changes

SCAN is operationalized for clinical curriculum design, supervision, and Assessment. The shift from misuse to misclassification is not merely semantic: it gives educators a concrete perspective on what to look for, what to assess, and what to intervene on. The framework grounds AI tutors in a theory of learning rather than treating them as generic productivity tools, and it connects Cognitive Offloading concerns to the developmental state of the individual learner.

What this means for practice

• Instructors. Have learners classify a task into a SCAN mode before they open a Generative AI tool, and hold Non-Negotiable tasks to unaided performance so that foundational competence is formed rather than delegated.
• Instructors. Read passive engagement inside a correctly classified AI task as a mis-skilling signal and shift the learner from AI assistance to expert assistance, with the supervising clinician acting as an epistemic auditor.
• Designers. Classify by where the individual learner stands relative to the task rather than by training phase, since the same task can be Aid for a first-year learner and Substitute for a resident.
• Instructors. Assess clinical reasoning through process rather than output: outcome measures cannot distinguish a correct answer produced from appropriate subzone classification from one produced by misclassification.
• Designers. Build supervisor training and metacognitive debriefing protocols into the curriculum, because SCAN's implementation depends on process-oriented supervision that is more demanding than the output-focused practice that currently predominates.

Limitations

• SCAN remains empirically untested: the authors state that its central constructs — subzone classification accuracy, engagement quality, and the mis-skilling pathway — await the validation studies outlined in their research agenda.
• The framework requires metacognitive capacity that novices may not possess: first-year students asked to classify tasks into SCAN zones may lack the domain knowledge to judge what a task genuinely requires, consistent with evidence that metacognitive accuracy develops with domain expertise (Kruger & Dunning, 1999).
• Never-skilling is not an established phenomenon; the authors state that it requires direct clinical empirical evidence in the future.
• Implementation depends on supervisor training and on institutional willingness to invest in process-oriented debriefing, and the authors note the paper reflects a fast-moving intersection of medical education and AI, so its claims should be read as theoretically grounded hypotheses rather than findings.

Citation

Tsim, F., Gutoreva, A., Weiss, A., & Dubosh, N. (2026). AI as Teammate: Rethinking Task Distribution in Medical Training. arXiv:2608.28373.