New paper argues AI explanations must be conversational, not static

In a new provocation paper accepted at the ACM Conversational User Interfaces (CUI)'26 conference, researchers Niharika Mathur and Smit Desai challenge the current approach to explainable AI. They argue that as AI systems become increasingly conversational, a critical gap has emerged: explanations are still studied as static artifacts—text blocks or one-shot outputs—while in practice, they are experienced as dynamic, back-and-forth dialogues. The authors contend that the conversational layer around an explanation is not merely a delivery mechanism but a critical constituent of its effectiveness. Their work invites the HCI and CUI community to treat explanations as interactive exchanges shaped by factors like timing, tone, persona, and conversational history.

To illustrate their vision, Mathur and Desai introduce HC2XAI (Human-Centered Conversational XAI), a framework that repositions explanation as a co-constructed process between human and AI. They walk through three illustrative scenarios—such as a user questioning a recommendation or probing a decision—showing how a static explanation fails compared to one that adapts based on prior dialogue. The paper stops short of a full system, but its provocation is clear: the next frontier for explainability isn't better algorithms—it's better conversations. By embedding explanation design into conversational flow, AI assistants could become not only more transparent but also more trustworthy and usable in real-world settings.