RuleEdit: Failure-Guided Human-AI Model Editing with Prospective Impact Preview
Researchers have developed RuleEdit, a system designed to help humans and AI collaboratively edit AI models. RuleEdit identifies potential AI failures using interpretable signals from rule tables and allows users to preview the impact of their edits on performance and embedding shifts. In a study involving health professionals, RuleEdit improved human-AI performance by over 14% and significantly enhanced the effectiveness of user feedback for model adaptation. AI
IMPACT Enhances human-AI collaboration in model development, potentially leading to more robust and controllable AI systems.