Neonatologists want AI as a second set of eyes—not the final word
Seventeen clinicians at one Dutch academic hospital welcomed AI for monitoring fragile newborns, but tied trust to local validation, human override and skills that must survive automation. The interviews reveal conditions for adoption, not evidence that an AI tool improved care.
The short version
- What happened
- The study interviewed 17 neonatologists and trainees at one Dutch academic centre where two neonatal AI systems were being piloted.
- Who it affects
- For babies and families, the immediate issue is not whether an algorithm sounds trustworthy but whether it has been validated for similar patients, whether clinicians can challenge it and who remains accountable when signals conflict.
- What remains unproven
- One primary source; 17 participants at one AI-experienced Dutch academic hospital; qualitative self-report; no diagnostic-accuracy comparator, patient outcomes or observed bedside decisions; final edited Version of Record is pending.
Keep in mind
One primary source; 17 participants at one AI-experienced Dutch academic hospital; qualitative self-report; no diagnostic-accuracy comparator, patient outcomes or observed bedside decisions; final edited Version of Record is pending.
Source analysis · 1 linked source. Primary analysed source date: 17 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
Trust was conditional on keeping the clinician in control
The clearest finding is not that neonatologists trust AI. It is that they would consider relying on it only under conditions that preserve clinical control. Researchers interviewed 17 neonatologists and trainees at University Medical Center Utrecht, where teams had piloted one system for early detection of late-onset sepsis and another using electroencephalography to predict outcomes for extremely premature infants. Participants pictured AI as a continuous monitor or alarm: something that could filter more data than a person, flag a pattern and prompt a bedside assessment.
That role stopped short of automated authority. Clinicians wanted to know when they could follow or disregard a recommendation, retain the ability to intervene and compare an output with a baby's responsiveness, colour, family context and their own clinical judgement. Several worried that a validated alert could still create false reassurance or alarm fatigue. Others raised the opposite risk: legal or organisational pressure to follow an algorithm even when their assessment disagreed. The paper therefore maps acceptable reliance rather than a simple level of trust.[1]
The evidence came from detailed interviews, not a deployment trial
Twenty-six eligible clinicians were invited and 17 participated: nine women and eight men, with experience ranging from under one year to more than 21 years. Interviews lasted 29 to 59 minutes between November 2023 and May 2024, in person or online, and continued until the researchers judged that meaning saturation had been reached. The team analysed Dutch transcripts inductively in NVivo, checked two transcripts between coders and organised the results around trusting AI in practice, maintaining control and limits of trust.
This design can explain how clinicians reason, but it cannot tell us whether they would behave the same way under pressure or whether any system is safe. Everyone worked at the same specialised hospital and had at least some exposure to neonatal AI pilots. The researchers did not record each participant's level of AI expertise, and translating selected quotations from Dutch may have lost nuance. Parents, nurses and patients were outside the sample. The full transcripts are unavailable because they contain sensitive data.[1]
The distinction that matters
A clinician's proposed trust path
The interviews describe a sequence for justified reliance, not a tested safety protocol.
Validate
Test on the local patient population
Know where performance holds
Interpret
Combine the alert with bedside evidence
Check whether it fits this baby
Override
Keep a trained clinician able to intervene
Human responsibility remains
For hospitals, validation must be local and continuing
Participants did not treat regulatory approval or a headline accuracy score as sufficient. They wanted evidence that a tool had been tested in a population resembling their own patients, clarity about which variables shaped validation, repeated performance in practice and training that would help them interpret probabilities such as a sepsis risk. They also wanted conventional clinical skills preserved for exceptions, outages and cases where an algorithm's data pattern conflicts with the whole patient.
The practical implication is that procurement should not end with buying a model. A hospital would need a local validation plan, explicit override rules, monitoring for alert fatigue and performance drift, education for every clinical role and a route for reviewing disagreements between system and clinician. Families also need a voice because decisions concern especially vulnerable patients. Confidence would rise through multicentre studies—including non-academic and lower-resource neonatal units—and prospective observation of decisions and patient outcomes. Until then, this single-centre study supports a design principle: AI may extend attention, but responsibility and contextual judgement stay human.[1]