When the algorithm enters the classroom: A critical integrative review of large language models, nursing education structural gaps, and the reconstitution of professional identity

Synthesis: Sun et al. (2026) argue that large language model (LLM) research in nursing education has asked what LLMs can do while neglecting what their integration does to the processes through which a novice becomes an expert nurse. This critical integrative review of 489 studies across 47 countries reframes the evidence through an integrated lens (Benner's skill acquisition, cognitive load theory, automation bias, Wenger's identity formation), finding the same technology both enhances and erodes nursing competence depending on whether it displaces cognitive work that is extraneous to, or constitutive of, the competence being developed. The authors' Evidence Gap Map shows the domains of greatest policy consequence (professional identity, relational and ethical competency, long-term outcomes) rest on the least rigorous evidence, and they introduce the Professional Identity Tension Model and Structural Empathy Suppression to guide future research and design.

Key Findings

1. LLMs address real structural gaps, but benefits are consistently moderated by context. LLMs help fill deficits in individualized learning support at scale, uneven access to clinical Simulation, and weak Scaffolding for progressive professional competency. Yet gains depend on the implementation framework (AI that preserves student cognitive agency, e.g. a nurse–AI collaboration RCT beating nurse-only care on patient satisfaction), on economic equity (paid-tier access barriers), and on model accuracy (RAG (Retrieval-Augmented Generation)-customized systems lifted nursing-specific accuracy from 43.27% to 73.29%).
2. Competency enhancement vs. competency attrition. Displacing extraneous cognitive load (documentation, routine retrieval) frees working memory for intrinsic processing; displacing intrinsic load (a full reasoning chain, an ethical justification, an individualized care plan) removes the object of learning itself. Unstructured AI reliance was linked to measurable deficits: students using ChatGPT as a sole resource scored significantly below textbook controls on ethical standards and clinical reasoning, and an AI-integrated curriculum produced higher scores yet less individualized, weaker-logic care plans — a performance–learning dissociation.
3. The evidence base is inverted. The Evidence Gap Map found no randomized or quasi-experimental studies of professional identity or career development, only three quasi-experimental studies of relational and ethical competency, and no study with follow-up beyond 12 months, while controlled evidence concentrates in cognitive and technical outcomes. Opinion/conceptual evidence dominated the highest-stakes domains, and over 50% of studies came from just three countries.

The Professional Identity Tension Model

The review's central conceptual contribution formalizes a single criterion — whether LLM-displaced cognitive and participatory work is extraneous to or constitutive of the competence being developed — across three levels: the task layer (which cognitive activities are displaced and whether they served a developmental function), the competency layer (enhancement when extraneous demands are displaced with cognitive agency preserved, vs. attrition when constitutive reasoning is displaced), and the identity layer (whether LLMs substitute for the participatory experiences through which professional identity forms). The design principle that follows: integration that offloads extraneous demand while preserving constructive or interactive engagement supports professional development; integration that displaces work constitutive of competence silently substitutes for it.

Structural Empathy Suppression

The authors propose this mechanism to reframe findings where AI appears to outperform nurses on relational metrics (e.g. an AI scoring higher than nurses on empathy ratings in a high-volume, 54.9 cases/hour context). Distinct from compassion fatigue (individual exhaustion) and emotional labor (managing emotional expression), it operates structurally: when systemic overwork makes authentic human empathic expression unsustainable and algorithmically consistent AI responses fill the gap, the AI appears comparatively empathetic — and the identity-constituting significance of nursing practice transfers to the algorithm. It reframes the policy question from "how effective is this technology?" to "what conditions are making this technology appear necessary?"

What this means for practice

• Administrators. Investigate the structural conditions that make AI look like the answer before adopting it: the review's Structural Empathy Suppression mechanism suggests an agent can outscore nurses on empathy ratings when caseloads and staffing already make sustained human empathy unsustainable.
• Curriculum designers. Keep participatory formation experiences — clinical simulation, placement, patient and peer interaction — required and non-substitutable, because professional identity forms through participation rather than through receiving correct information.
• Researchers. Target the highest-consequence gaps: the Evidence Gap Map found zero randomized or quasi-experimental studies of professional identity, only three of relational and ethical competency, and no study with follow-up beyond 12 months — and measure capability rather than performance, since outcomes measured while an LLM is available may index fluent performance rather than durable capability, so use delayed, no-tool post-tests and transfer tasks on unfamiliar clinical presentations.
• Policymakers. Require every implementation framework to state which cognitive work the LLM displaces and whether that work was extraneous to or constitutive of the competence being developed, turning the review's dual-pathway criterion into procurement and approval criteria — and require evidence that implementations preserve access as well as developmental processes, since paid-tier subscription costs and detection-tool bias against non-native speakers can widen inequities.
• Medical educators. Build structured nurse–AI collaboration exercises into clinical courses rather than AI answer services — position LLMs as reasoning partners rather than answer sources, and judge them by their effect on students' independent reasoning, judgment and reflection rather than by workload reduction — since controlled evidence favors implementations that preserve student cognitive agency.

Limitations

• The search was restricted to English-language publications and omitted the theoretical frameworks used for analysis as search terms, so studies engaging those frameworks may have surfaced only incidentally, and screening agreement was moderate-to-substantial (κ = 0.67 and κ = 0.78), leaving edge-case bias unexcluded.
• The Professional Identity Tension Model is preliminary and has not been empirically tested; Structural Empathy Suppression rests primarily on a single RCT (Wan et al., 2024), and one source informing the regulatory implications is a preprint pending peer review.
• Longitudinal evidence is near-absent, so conclusions about professional identity formation are inferred from cross-sectional and short-term data, and over 50% of included studies came from just three countries.

Citation

Sun, Y., Li, H., Tao, X., Zhou, X., Gururajan, R., & Zhang, J. (2026). When the algorithm enters the classroom: A critical integrative review of large language models, nursing education structural gaps, and the reconstitution of professional identity. Computers and Education: Artificial Intelligence, 11, 100675.