Radiologists Showed Automation Bias in AI Bone Studies
A study found that junior radiologists were more susceptible to incorrect AI guidance when estimating bone age.
Updated on Sept. 28, 2026 in Artificial Intelligence
Live Poll
Do you trust artificial intelligence outputs when you receive medical diagnoses from a doctor?
Researchers published a randomized crossover study showing that five out of six radiologists performed differently when presented with accurate versus sham AI-assisted bone age assessments. This research highlights the influence of automation bias on clinical diagnostic workflows.
Why it matters
Understanding how clinicians interact with diagnostic tools is critical as AI integration accelerates in medical imaging. The findings suggest that professional experience levels significantly alter how radiologists process machine-generated recommendations.
Participants assessed 200 radiographs using both true and sham AI outputs; the most accurate radiologist showed a p-value of 0.297, while the other five participants recorded a p-value of 0.001. Errors increased significantly when the AI discrepancy exceeded six months.
The players
Nature
A leading peer-reviewed scientific journal that publishes fundamental research across physical and life sciences.
The details
Researchers employed paired t-tests to evaluate the mean absolute difference in bone age assessments between accurate AI and sham AI conditions, where sham outputs were generated by randomizing true-AI predictions. The study revealed that senior radiologists were more likely to disagree with algorithmic inputs, while junior participants demonstrated higher susceptibility to automation bias—a cognitive tendency to favor suggestions from automated systems regardless of accuracy.
Timeline
September 28, 2026: The research findings were published.
The Tech Race
This study adds to the growing literature on human-in-the-loop diagnostic benchmarks where accuracy alone is increasingly viewed as insufficient for safe deployment. It contrasts with standard technical benchmarks that evaluate algorithmic performance in isolation from clinical oversight.
Medical facilities utilizing AI-assisted imaging tools may need to implement training protocols that explicitly address automation bias. Diagnostic workflows could evolve to require independent verification by senior staff when AI-provided discrepancies exceed the six-month threshold.
The takeaway
The study confirms that experience level remains a primary defense against algorithmic bias in diagnostic settings. Stakeholders should look for upcoming larger-scale validation studies to confirm if junior radiologist susceptibility can be mitigated through interface design.
Further reading
For broader trends in medical software testing, explore the latest Artificial Intelligence research.
More information
View the complete findings in the peer-reviewed research article.
Live Poll
Do you trust artificial intelligence outputs when you receive medical diagnoses from a doctor?
