Editable XAI: Toward Bidirectional Human-AI Alignment with Co-Editable Explanations of Interpretable Attributes

Users can now actively shape AI explanations rather than passively read them. The study found that allowing users to edit explanations—writing their own rules about how the AI should work—improved both user understanding and alignment between human and AI decision-making compared to traditional read-only explanations.

Editable explanations work better than static ones for learning. CoExplain required fewer user edits and less time to achieve alignment compared to manually editable alternatives, suggesting that well-designed interactive systems can make AI more understandable and controllable without overwhelming users.

The system bridges neural networks and human reasoning. CoExplain translates complex neural network decisions into a decision tree that humans can understand, then converts user-written rules back into the AI system—creating a genuine two-way conversation between humans and AI rather than one-directional explanation.

Domain knowledge mismatches are a real problem in AI deployment. The paper identifies that disagreements between what users know about their field and what the AI "thinks" it knows creates confusion and reduces trust—a practical challenge that editable explanations directly address.

Active participation improves comprehension. The user study (N=43) demonstrated that the act of writing rules to edit explanations itself helped users understand the AI better, suggesting that responsible AI systems should invite user participation rather than just provide information.