The distinction at the heart of the study is what the authors call mechanism-aware benchmarking, and it matters because the two error types carry different dangers. A true hallucination occurs when a model invents a contraindication that does not exist, for example claiming a drug cannot be used in a situation where it is actually the recommended choice. A false contraindication, by contrast, is a real-world clinical caution that the model applies incorrectly, refusing to recommend an appropriate intervention because it misreads the context. Both can lead a trainee astray, but they stem from different failure mechanisms inside the model, and separating them allows educators to understand precisely where and why these systems go wrong rather than simply tallying a single error score.

To conduct the evaluation, the researchers presented fifty cases of pediatric difficult airway management to five large language models using a standardized prompt anchored in the 2022 American Society of Anesthesiologists Difficult Airway Guidelines. Two anesthesiologists then rated the responses on a modified 0 to 5 scale in a blinded, comparative assessment. The reliability of that human scoring was extraordinary: the inter-rater agreement, measured by a quadratic weighted kappa, reached 0.982, a value that essentially approaches perfect concordance and lends considerable statistical weight to the findings. The differences between the models themselves were also highly significant, with a Friedman test returning a chi-squared value of 100.12 across four degrees of freedom and a p-value below 0.001, confirming that the performance gaps were not statistical noise.

In this single-pass evaluation, GPT-5.2 Thinking emerged as the strongest performer, achieving both the highest mean score and the largest proportion of responses judged acceptable. At the other end of the spectrum, Claude Opus 4.5 recorded the highest combined critical-error rate at 23.0 percent, compared with 6.5 percent for both GPT models tested. Those numbers alone would be striking, but the mechanism-level analysis revealed something even more consequential: the models fail in characteristically different ways, and those failure signatures matter enormously when deciding whether such tools belong anywhere near a training environment.

True hallucinations, the fabrication of nonexistent contraindications, were overwhelmingly concentrated in one model. Claude Opus 4.5 produced them in 11.8 percent of cases, while the other four models ranged from just 0.5 to 2.2 percent. False contraindications, however, proved to be a universal weakness, appearing across every model at rates between 5.0 and 13.0 percent. Perhaps most tellingly, these false contraindications clustered around a single drug: rocuronium, the neuromuscular blocking agent that is central to rapid sequence intubation and a cornerstone of emergency airway management. A model that hesitates to recommend rocuronium, or wrongly flags it as contraindicated, is not making a harmless stylistic error; it is steering a learner away from a potentially lifesaving intervention in the very scenarios where seconds count.

The study also uncovered a gradient of failure that tracks directly with clinical complexity. The proportion of responses judged inadequate, defined as a score of 2 or below, rose steadily as scenarios became harder. For difficult intubation cases, inadequacy ranged from 32 to 98 percent across the models. For difficult ventilation, it climbed to between 46 and 94 percent. And for the most dire scenario of all, cannot intubate cannot oxygenate, known in the field as CICO, inadequacy spanned 52 to 92 percent. In other words, the situations where a trainee most needs accurate, guideline-concordant guidance are precisely the situations where these models are most likely to fall short, a pattern that inverts the usual assumption that AI assistance is most valuable in the hardest cases.

The authors’ conclusion is carefully calibrated but unambiguous. Contemporary LLMs exhibit model-specific performance characteristics and error patterns in pediatric difficult airway management, and dosing-related failures manifest through distinct mechanisms of true hallucination versus false contraindication. The findings reveal what the researchers describe as a marked inadequacy of the scenarios employed, particularly as complexity and critical error rates increase. Their verdict on deployment is equally measured: current LLMs may hold value as supervised supplementary learning tools, but they should not function as autonomous sources of clinical or educational guidance. That framing positions these systems as something closer to a study partner whose answers must always be checked, rather than a reference whose word can be trusted.

Why does pediatric difficult airway management stress these models so severely? The clinical domain itself offers clues. Children are not small adults; airway anatomy, drug dosing, and equipment sizing all scale with age and weight in ways that demand precise, patient-specific calculation. The ASA’s 2022 difficult airway guidelines embed a structured decision tree that models must navigate step by step, and any drift at an early node cascades into a wrong endpoint. Dosing errors are especially hazardous in this population because the therapeutic window for neuromuscular blockers and induction agents is narrow, and the study’s finding that errors concentrate around rocuronium suggests the models struggle most where weight-based calculation intersects with urgency and contraindication logic.

The study’s methodology also deserves attention as a template for future AI evaluation in medicine. Rather than asking whether a model’s answer is simply right or wrong, the mechanism-aware approach asks what kind of wrong it is, and that granularity has practical consequences. An educator deploying an LLM as a teaching aid can now anticipate that one model may invent contraindications out of thin air while another may reflexively withhold appropriate drugs, and can design supervision and debriefing around those known tendencies. The near-perfect inter-rater agreement achieved by the two blinded anesthesiologist raters further demonstrates that expert human judgment can reliably and reproducibly grade the quality of AI clinical reasoning, providing a credible benchmarking standard.

For the broader conversation about artificial intelligence in medicine, the study lands at a moment of intense enthusiasm and equally intense anxiety. It neither condemns LLMs outright nor licenses their casual use; instead it draws a precise boundary. As informal just-in-time learning tools for trainees, under the eye of an experienced supervisor, these models may stimulate reasoning and provide a starting point for discussion. As autonomous advisors in pediatric airway emergencies, they are not ready, and the data show why: critical error rates as high as 23 percent, universal false contraindications around essential drugs, and inadequacy rates that spike exactly when stakes peak. The message for medical educators, and for the developers of these systems, is that safety in high-stakes clinical education must be demonstrated, not assumed, and that the next generation of benchmarks should measure not just whether models answer, but how they fail.

Subject of Research: Benchmarking large language models for safety and error mechanisms in pediatric difficult airway medical training

Article Title: Mechanism-aware benchmarking of large language models as learning aids in pediatric difficult airway training: true hallucinations vs. false contraindications

Article References: Yediyıldız, M. B., & Durmuş, İ. (2026). Mechanism-aware benchmarking of large language models as learning aids in pediatric difficult airway training: true hallucinations vs. false contraindications. BMC Medical Education. https://doi.org/10.1186/s12909-026-10509-y

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10509-y

Keywords: large language models, pediatric anesthesia, difficult airway, hallucination, false contraindication, rocuronium, ASA difficult airway guidelines, medical education, simulation-based training, drug dosing, CICO, artificial intelligence