Presentation
The Influence of Positive and Negative Framings of AI's Capabilities on the Effectiveness of AI Explainability
DescriptionExplainable artificial intelligence (XAI) is proposed to help humans develop appropriate dependence on AI by providing information about cases that AI performs almost perfectly on (non-error-prone cases) and cases that AI might make mistakes (error-prone cases). However, this information contains both positive and negative aspects of AI performance, which might affect decision-making differently. Thus, the current study investigated how different valences in XAI framings regarding AI's performance affected dependence and subjective outcomes, especially with the rarer occurrence of error-prone cases. In an online experiment, participants (N=176) estimated the percentage of bacterial contamination of images and were informed that AI would support this task by providing recommendations. Additionally, participants in some groups were informed of AI's strengths and/or weaknesses. The results showed that providing information about AI's error-proneness might not always be beneficial for appropriate behavioral adaptation. Due to a negativity bias, positive information about the AI's strength might not be utilized when the information implies a possibility for negative aspects of AI performance. This suggests that XAI should be carefully designed and implemented while considering potential cognitive biases and task characteristics to mitigate the negative effects of XAI on dependence.
Event Type
Lecture
TimeWednesday, October 15th2:10pm - 2:30pm CDT
LocationGrand B
Human AI Robot Teaming (AI)
Similar Presentations