Radiology Diagnostic Integration and Human Underweighting
Diagnostic Classification Accuracy moved from 82.5% to 79.8% — a +-2.7 percentage-point (-3% relative) change (sample: 200 radiologists, 10,000 cases).
Study details
- Challenge
- Radiologists interpreting chest X-rays have a high base error rate but resist relying on computer-aided diagnostic recommendations, which leads to suboptimal patient outcomes.
- Intervention
- Radiologists were given access to an AI pathology detector displaying exact probability scores (0-100%) for critical conditions like pneumonia, but were otherwise left to make final determinations.
- Outcome
- Diagnostic Classification Accuracy: 82.5% (control) → 79.8% (treatment); +-2.7 pts, -3% relative
- Sample
- 200 radiologists, 10,000 cases
- Key takeaway
- Joint decision-making underperformed AI working alone (85.2% accuracy) because humans underweighted the AI output and treated it as independent rather than correlated information, exposing a core vulnerability in collaborative choice architecture.
- Biases leveraged
- Algorithm AversionCorrelation Neglect
- Source
- Agarwal, N., Moehring, A., Rajpurkar, P., & Salz, T. (2023). Combining human expertise with artificial intelligence: Experimental evidence from radiology. NBER Working Paper No. 31422. ↗ source
Cite this page
Behavioral Economics Lab. "Radiology Diagnostic Integration and Human Underweighting." Behavioral Economics Lab, https://behavioraleconomicslab.com/findings/radiology-ai-diagnostic.
Primary source: Agarwal, N., Moehring, A., Rajpurkar, P., & Salz, T. (2023). Combining human expertise with artificial intelligence: Experimental evidence from radiology. NBER Working Paper No. 31422.