Artificial Intelligence and the Future of Bedside Nursing: A Critical Narrative Review of Augmentation, Substitution and the Human-Centred Artificial Intelligence Partnership Framework

Abstract

Background: Artificial intelligence is entering acute and long-term care at a pace that has outstripped its evidence base, and commentary frequently frames the change as a contest between machines and the nursing workforce. Nursing carries the surveillance function on which timely rescue of deteriorating patients depends, so whether computational systems substitute for or extend that function has substantial policy consequences. Purpose and Scope: This critical narrative review evaluates the strength, consistency and methodological quality of evidence bearing on the substitution and augmentation hypotheses in bedside nursing, restricted to registered nurses delivering direct care in hospital, residential and home settings. Educational applications, nursing management analytics and consumer-facing health applications were excluded. Approach: Literature was identified through open scholarly databases and indexes, citation searching and targeted institutional sources, appraised for design adequacy, validation status, generalisability and reporting transparency, and synthesised thematically around mechanisms, controversies and methodological problems rather than as a study-by-study catalogue. Principal Findings: Evidence is strongest where computational prediction has been embedded in a nurse-delivered response pathway, and weakest where models have been evaluated in isolation from the clinical work that gives them effect. Widely deployed proprietary sepsis models have shown poor discrimination on external validation and have often alerted only after clinicians had already acted, whereas deterioration programmes coupling automated risk scores to nurse review and rapid-response escalation have been associated with reduced mortality in large non-randomised evaluations. Documented nursing judgement has demonstrated predictive value that physiological models do not replicate. Documentation, virtual nursing and robotic applications show efficiency signals accompanied by unresolved problems of note quality, workload redistribution and measurement heterogeneity, and automation bias, deskilling, algorithmic inequity and unmeasured work transfer are documented rather than speculative risks. Implications: The available evidence supports a partnership model in which computational systems assume selected informational and monitoring tasks under nursing authority, and does not support substitution of the bedside role. A Human-Centred Artificial Intelligence Partnership framework is proposed, comprising five domains with explicitly labelled established and hypothesised relationships, to structure evaluation and to make partnership claims falsifiable rather than aspirational.