Increasing transparency of computer-aided detection impairs decision-making in visual search

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Human-computer interaction

Background:

• Artificial intelligence (AI) is increasingly used in healthcare, prompting calls for transparency in AI systems.
• Transparency recommendations aim to inform users about AI accuracy and functionality.
• However, enhanced transparency may lead to overreliance and negative outcomes in human decision-making.

Purpose of the Study:

• To investigate the impact of AI transparency on human decision-making in a medical screening context.
• To assess how knowledge of AI accuracy affects performance in a visual search task.

Main Methods:

• A simulated laboratory mammography task was used, involving visual search for cancer.
• Participants' performance was evaluated under two conditions: transparent (AI accuracy disclosed) and non-transparent (AI accuracy withheld).
• Computer-aided detection (CAD) systems with varying accuracies provided AI prompts.

Main Results:

• Increased AI transparency impaired task performance.
• The transparent condition resulted in more false alarms and decreased sensitivity.
• Recall rate increased, and positive predictive value decreased with greater transparency.

Conclusions:

• Transparency in AI systems, particularly in medical screening, can negatively affect human decision-making.
• Overtrust in AI due to increased transparency may lead to adverse clinical outcomes.
• Further research is crucial to understand the complex relationship between AI transparency and human performance in healthcare settings.