Large language models as neurosurgical clinical assistants: evidence, failure modes, and a hypothesis-generating framework for task-level evaluation

Neurosurgical evaluation of large language models (LLMs) has focused largely on examination performance, which does not establish clinical safety or workflow value. We conducted a structured clinical narrative review of PubMed/MEDLINE and publisher databases (May–September 2026, updated October 2026) across five tasks: treatment planning, clinical documentation, patient communication, literature synthesis, and clinical reasoning, with additional consideration of hallucination and medicolegal issues. Prospective studies, randomized trials, systematic reviews, and studies using real patient data were prioritized. Evidence quality was heterogeneous and largely single-centre. AI-drafted neurosurgical operative notes showed comparable accuracy but lower content (comprehensiveness) ratings than surgeon-written notes. In a prospective neuro-oncology study, GPT-4 agreed with a tumour board in 76% of 101 cases versus 90% for a specialist; most discordances were recommendations for treatment the board declined, whereas all discordances against guidelines in low-complexity cases were omissions of indicated radiotherapy. Reference-generation studies showed substantial fabrication, while title and abstract screening achieved high sensitivity at low precision. Randomized reasoning studies did not show reliable benefit from model assistance. Recent neurosurgical studies using real patient data were small pilots, and none measured patient-level outcomes. Current evidence supports task-specific evaluation rather than global judgments of LLM capability. We propose a consequence–verifiability matrix and staged evaluation pathway as expert-derived hypotheses for prospective testing, not validated clinical governance instruments.