Can Artificial Intelligence Distinguish Helpful Complexity from Harmful Polypharmacy? A Critical Evidence Review

Polypharmacy may represent necessary treatment for multimorbidity or expose patients to avoidable harm. Artificial intelligence is proposed to identify inappropriate medicines, predict interactions, and support deprescribing, but model accuracy does not establish whether a regimen’s complexity is beneficial or harmful. To critically examine whether current systems distinguish helpful medication complexity from harmful polypharmacy, and to identify the evidence required for that claim. A bounded critical evidence review used seven frozen Google Scholar searches, supplementary citation searching, and publisher, PubMed, and Crossref verification. Of 104 captured records, 28 duplicates were removed, 76 were screened, 45 reports were sought, 43 were assessed, and 21 reports representing 21 studies or reviews were included. Evidence was extracted by construct, target, reference standard, validation design, patient context, clinical utility, and risk of inappropriate simplification. Prediction studies, interventions, reviews, and contextual studies were appraised using appropriate domains and synthesized without pooling. Most systems predicted criteria-defined potentially inappropriate medication labels, compiled rule-based warnings, estimated interaction or hospitalization risk, or generated clinician-mediated deprescribing suggestions. Prospective systems could increase discontinuation or reduce medication burden, yet clinical-outcome findings were limited and mixed. Patient goals, prognosis, indication, time-to-benefit, and treatment burden were rarely represented directly. Proxy-label circularity, restricted external validation, and outcome disconnect limited interpretation. Current evidence supports artificial intelligence as a source of partial medication-risk signals, not as a demonstrated judge of patient-specific regimen value. Distinguishing helpful complexity from harm requires explicit benefit representation, contextual goals, counterfactual comparison, prospective evaluation, human adjudication, and safety monitoring.