The mirage of consensus: rethinking AI-driven delphi simulations in surgical expert panels – letter to the editor

To the Editor, The recent study by Park et al[1] represents a remarkable step toward understanding how large language models (LLMs) might emulate the iterative consensus-building of Delphi panels. By translating expert deliberation into algorithmic dialogue, the authors gesture toward a future in which medical consensus could be accelerated, standardized, and – perhaps – automated. Notably, the study complies with the TITAN 2025 guidelines for transparent and ethical reporting of AI-assisted research[2], further underscoring its methodological ambition. Yet beneath this technical triumph lies a deeper epistemic tension: when does consensus cease to be knowledge and become illusion? The study’s design, while methodologically meticulous, positions artificial intelligence as a self-contained epistemic collective, effectively simulating expert panels without engaging with them. Yet, the most plausible future of surgical deliberation is not one where AI supplants human judgment, but one where it integrates into it – as a cognitive catalyst rather than a sovereign voice[3]. In omitting a hybrid human-AI control group, the study inadvertently bypasses the central question: does AI enrich human consensus-building or merely echo its statistical patterns[4]? The experiment models substitution when the more urgent challenge lies in optimizing collaboration. Equally thought provoking is the study’s elevation of consensus rate as the definitive performance metric. Consensus, however elegantly attained, does not inherently confer validity. Surgical history is replete with instances of collectively endorsed practices later refuted by evidence. In this context, the high agreement ratios achieved by LLMs may signal syntactic coherence rather than semantic comprehension. To conflate linguistic symmetry with clinical soundness is to risk enshrining fluency over fidelity. Robust evaluation demands more than internal alignment; it requires blinded expert adjudication, or better still, triangulation with established clinical guidelines, to determine whether AI-generated consensus possesses epistemic integrity or merely rhetorical fluency.HIGHLIGHTS AI-simulated Delphi panels risk replacing dialogue with statistical imitation. Consensus without human input may mirror fluency, not clinical soundness. Study lacks hybrid human–AI controls to test true deliberative enhancement. High AI agreement may mask conflict avoidance, not epistemic integrity. Disagreement zones are key to clinical judgment, not failures of consensus. Indeed, the most telling aspects of the simulation may lie not in what the AI agrees upon, but in what it leaves unresolved. The study offers little reflection on the “non-consensus” items – the very liminal spaces where human experts grapple with ethical ambiguity, evidentiary uncertainty, and therapeutic pluralism. These contested zones are not failures of consensus but sites of epistemic richness, where meaningful deliberation occurs. By eliding these domains, the simulation treats consensus as a terminal output rather than an iterative process, thereby overlooking the dialectical role of disagreement in refining clinical judgment. Compounding this illusion is the nature of language itself. LLMs are not optimized to interrogate or dissent, but to harmonize – to hedge, soften, and converge toward the statistical center of discourse. The appearance of unanimity may thus reflect not true accord, but an algorithmic aversion to conflict[5]. This phenomenon risks creating a “consensus mirage,” where rhetorical smoothness conceals unresolved controversies. In surgical science – where diagnostic ambiguity and therapeutic diversity are normative rather than exceptional – such flattening is not benign[6]. Employing linguistic analytics, such as hedging indices or semantic centrality metrics, could help distinguish between authentic deliberative synthesis and superficial textual convergence. In summary, while the study admirably advances the technical frontiers of AI-assisted consensus modeling, it simultaneously prompts deeper reflection on the epistemic risks of artificial agreement. What matters in surgical decision-making is not merely that agents reach consensus, but whether that consensus bears the weight of truth, withstands clinical scrutiny, and respects the pluralism of medical realities. The future lies not in asking whether machines can agree, but in ensuring that what they agree upon is worth believing. We urge future Delphi-AI designs to incorporate adversarial prompts, hybrid control arms, and real-world unresolved scenarios so that consensus is not only replicable but also epistemically grounded.