Expertise Alters How Clinicians Respond to AI in Neonatal Pneumoperitoneum
In the fastest-moving corners of neonatal care, a wrong read can push a newborn toward delayed surgery or unnecessary intervention. This study places neonatal pneumoperitoneum in that exact pressure zone and asks a more uncomfortable question than simple model accuracy: what happens when clinicians trust, resist, or overcorrect for machine advice?
Investigators tested human–AI collaboration in a multi-reader crossover design built around 1,750 interpretation events. Clinicians reviewed radiographs with help from either a highly reliable model or a deliberately error-injected model, creating a controlled setting in which diagnostic behavior could be observed rather than assumed. The central finding was not that AI helps or harms in a uniform way, but that expertise changes the direction of the effect. Trainees became more vulnerable to automation bias when the model appeared dependable, while the small specialist group showed a markedly different response when the algorithm erred.
One result stands out sharply: trainees accepted 52.0% of incorrect AI suggestions when the model was highly reliable.
A Reliable Model Can Still Train the Wrong Reflex
That pattern matters because confidence in an algorithm can become its own hazard. In high-stakes neonatal imaging, a seemingly accurate system may teach less experienced clinicians to defer too quickly, especially when the answer appears to arrive with machine authority. The diagnosis at stake is not an academic label; pneumoperitoneum in a neonate can represent surgical emergency, where minutes and confidence both carry clinical cost.
The study’s framing around “never-skilling” and deskilling is clinically plausible. If trainees repeatedly encounter correct machine outputs, they may become faster at agreement than at independent pattern recognition. Once that habit forms, the system can look safe on validation metrics while quietly eroding the clinician’s ability to detect the rare but dangerous miss. In this setting, a dependable model does not automatically produce a dependable team.
Specialists Behaved More Like Sentinels Than Followers
When the error-injected model was placed before expert neonatologists, their behavior shifted in the opposite direction. The three participating neonatologists correctly overrode 91.7% of errors, with a Wilson 95% confidence interval of 83.0% to 96.0%. The investigators describe this as a sentinel behavior phenotype, and the term fits the observed dynamic: heightened scrutiny, active challenge of the algorithm, and a return to deliberate clinical reasoning when the machine faltered.
The sample is too small to convert that observation into a general rule, and the study itself treats it as hypothesis-generating. Even so, the bedside implication is hard to ignore. Expert clinicians may not simply “use AI better”; they may use it differently, with an instinct to interrogate discordant outputs rather than absorb them. In emergency neonatal radiology, that form of adversarial vigilance can be a protective skill rather than a nuisance.
This creates a striking asymmetry across training levels. A highly reliable system tempted trainees toward passive acceptance, while flawed AI seemed to activate a corrective stance in experts. The same tool, in other words, did not merely assist diagnosis; it re-shaped diagnostic posture.
Why System Resilience Matters More Than Standalone Accuracy
The paper’s most important conceptual move is to separate model performance from team performance. Standalone accuracy can look impressive in isolation, yet clinical validity depends on whether the human–AI pair remains reliable when the algorithm fails, misfires, or drifts outside its training comfort zone. In a time-critical neonatal emergency, resilience is not a theoretical metric. It is the capacity to preserve diagnostic integrity under pressure, even when the interface presents a false sense of certainty.
That distinction has direct implications for how AI is introduced into neonatal workflows. Systems designed only to maximize correct outputs may still leave clinicians exposed to automation bias, especially those early in training. An AI tool for acute neonatal imaging must therefore be judged not only by discrimination statistics, but also by its effect on human judgment, error detection, and override behavior. The risk is not just that a model makes mistakes; it is that the surrounding team stops noticing them.
The study also gives concrete shape to a wider concern in digital medicine: educational exposure can either inoculate or condition clinicians. If a sandbox can deliberately surface algorithmic failure, it may help clinicians practice skepticism before they encounter a real newborn whose abdomen contains free air and a narrow window for action. In that sense, simulation becomes more than pedagogy; it becomes risk control.
What the 1,750-Event Design Revealed About Human–AI Teams
The crossover structure gave the analysis an important advantage: the same broad clinical task could be seen under different algorithmic conditions. That makes the comparison less about individual preference and more about how the interaction itself changes behavior. The study is especially relevant because it focuses on a diagnosis that rewards speed but punishes complacency. Neonatal pneumoperitoneum is precisely the kind of case where a clinician must integrate image features, clinical context, and the possibility that the machine is confidently wrong.
The findings also suggest that expertise is not merely a buffer against AI error. It changes the entire strategy of decision-making. Experts appear more willing to pause, challenge, and override; trainees appear more likely to follow the automated lead when the model seems trustworthy. In practical terms, that means AI deployment should be tiered to user experience, not simply distributed uniformly across a service.
The authors release an open-source educational sandbox intended to expose clinicians to automated errors and reduce the risks of deskilling and never-skilling. The clinical rationale is straightforward: if people are going to work beside algorithms in emergency care, they need to practice seeing through algorithmic certainty before the real case arrives.
The study leaves one major issue unresolved: whether the sentinel behavior seen in three neonatologists will persist in a larger, more diverse specialist group under routine clinical conditions. That question will determine whether this is an idiosyncratic expert response or a reproducible safeguard for neonatal imaging teams.