ConvXAI: Delivering Heterogeneous AI Explanations via Conversations to Support Human-AI Scientific Writing

Conversational interfaces improve AI transparency: ConvXAI, a chatbot-based explanation system, outperformed traditional graphical interfaces in helping users understand AI writing assistance, suggesting that dialogue-based explanations may be more effective than static documentation for explaining AI decisions.

Users need multiple explanation types, not just one: The research identified "multifaceted" explanations as a key design principle, indicating that people require different types of information (such as reasoning, examples, and alternatives) to truly understand why an AI system made a particular suggestion—a finding that challenges one-size-fits-all explanation approaches.

Users should control the depth of explanations: The "controllability" design rationale emerged as important, meaning people want to decide how detailed an AI explanation should be rather than receiving fixed-length responses, allowing them to drill down into specifics only when needed.

Context matters for useful AI explanations: The "context-aware drill-down" principle showed that explanations are more helpful when they reference the specific document or writing task at hand, rather than providing generic information about how the AI works.

Practical AI systems need mixed-initiative interaction: The research emphasizes that effective human-AI collaboration requires both the user and the AI to take initiative in conversations, rather than one party simply responding to the other, reflecting how people naturally seek help and explanations.