Scholarlink

Human–AI collaboration in medical diagnostics offers significant potential, yet its effectiveness remains debated. This study investigates two key phenomena in hybrid intelligence systems: human augmentation, where AI improves user performance, and outperformance, where the human–AI team exceeds both individual components. In a multi-site study involving 330 doctors and 16,641 cases across six diagnostic modalities, participants consulted an AI with 81% accuracy during diagnostic tasks. Post-consultation accuracy increased from 75% to 79%, with 57% of participants improving. Outperformance occurred in 18% of cases overall and 27% among those initially less accurate than the AI. The largest gains were observed among lower performers, while 11% of participants—mostly those initially more accurate than the AI—saw declines. These results support the potential of well-integrated AI to enhance diagnostic accuracy and highlight the importance of fostering calibrated trust and effective interaction design to realize the benefits of hybrid intelligence while mitigating risks of over-reliance.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Who Knocks on Heaven’s Door: Measuring Augmentation and Outperformance in Human–AI Diagnostic Teams

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Federico Cabitza

摘要

Human–AI collaboration in medical diagnostics offers significant potential, yet its effectiveness remains debated. This study investigates two key phenomena in hybrid intelligence systems: human augmentation, where AI improves user performance, and outperformance, where the human–AI team exceeds both individual components. In a multi-site study involving 330 doctors and 16,641 cases across six diagnostic modalities, participants consulted an AI with 81% accuracy during diagnostic tasks. Post-consultation accuracy increased from 75% to 79%, with 57% of participants improving. Outperformance occurred in 18% of cases overall and 27% among those initially less accurate than the AI. The largest gains were observed among lower performers, while 11% of participants—mostly those initially more accurate than the AI—saw declines. These results support the potential of well-integrated AI to enhance diagnostic accuracy and highlight the importance of fostering calibrated trust and effective interaction design to realize the benefits of hybrid intelligence while mitigating risks of over-reliance.