The instrument is not the claim: LLMs and responsibility in radiology

Debates about large language models (LLMs) in publishing reprise longstanding concerns about externalized cognition.From Plato's Phaedrus to contemporary "extended mind" accounts, inscription and computation redistribute cognitive labor without eliminating human judgment.Radiology illustrates this lineage vividly: deep learning now contributes to reconstruction, segmentation, triage, and characterization, while language technologies shape how results are narrated.We argue for a radiology-specific boundary that aligns disclosure with evaluation.When AI systems participate in measurement or inference-altering data properties, producing quantitative features, steering workflow, or generating figures-they belong in Methods with sufficient specificity to permit scrutiny.When AI assists expression-clarifying prose, reducing jargon, or translating-its role is analogous to established editorial aids and does not, by itself, change the evidentiary object.This perspective does not reject disclosure; it redirects attention to what bears on claims.The central task remains constant: upholding human responsibility for study design, data integrity, and J o u r n a l P r e -p r o o f reasoning through transparent methods, auditable code and data where feasible, vigilant peer review, and openness to critique, while recognizing that LLMs can make sound radiological science more legible.