Traditional Explainable AI (XAI) often treats transparency as a "one-shot" delivery. A model produces a prediction, and an algorithm like SHAP or LIME generates a static heatmap or a list of feature importance scores. However, research increasingly shows that these "data dumps" often leave users—both experts and novices—with more questions than answers.
A growing body of work argues for a paradigm shift: treating explainability not as a static report, but as a collaborative, multi-turn dialogue. By moving from oracular outputs to interactive conversations, AI systems can align their internal logic with human mental models, resolve ambiguities through follow-up questions, and significantly improve user trust and task performance.
The Limits of Static Transparency
In high-stakes environments like cybersecurity or medicine, static visualizations often fail to provide the necessary context. Harry Li and colleagues in More Questions than Answers? found that traditional XAI tools can actually increase confusion for domain experts by failing to address specific diagnostic needs.
The core issue is that human explanation is fundamentally social. As Daniel D. Suthers, Beverly Park Woolf, and Matthew Cornell argued as early as 1992 in Steps from Explanation Planning to Model Construction Dialogues, effective explanation is a process of building a shared mental model. A static report cannot adapt to a user’s evolving understanding or handle shifts in conversational focus.
Explanation as a Bidirectional Loop
To bridge this gap, researchers are developing systems that support bidirectional communication. In these frameworks, the user is not a passive recipient but an active "inquisitor."
Systems like TalkToModel, developed by Dylan Slack, Himabindu Lakkaraju, and Sameer Singh, allow users to interrogate models using natural language. Their work in Explaining machine learning models with interactive natural language conversations using TalkToModel demonstrates that this conversational approach can nearly double the speed and accuracy of model analysis compared to traditional dashboards.
This bidirectional loop serves several critical functions:
- Resolving Ambiguity: Users can ask follow-up questions to clarify technical terms or specific model behaviors (Zhang et al., 2023).
- Counterfactual Exploration: Users can probe "what-if" scenarios to understand the boundaries of a model's decision logic (Kuzba et al., 2020).
- Managing Cognitive Load: Instead of overwhelming a user with all available data, systems like the "Mediators" proposed by Nils Feldhus and Sebastian Möller deliver information in small, digestible increments (Feldhus et al., 2022).
Formalizing the Interaction
The shift toward dialogue is not just an interface preference; it is being formalized through new computational frameworks.
David S. Watson and Luciano Floridi propose "The Explanation Game" (Watson & Floridi, 2021), which treats interpretability as a collaborative game between an information-seeking agent and a model-representing agent. Similarly, Antonio Rago, Francesca Toni, and colleagues advocate for Argumentative XAI (Rago et al., 2021). In this view, explanations are structured as dialectical debates where internal model components "support" or "attack" a conclusion, mirroring how humans weigh evidence.
Toward Agentic Interpretability
The most recent evolution in this field is Agentic Interpretability. Researchers like John Hewitt, Oyvind Tafjord, and Been Kim suggest that Large Language Models (LLMs) should act as proactive tutors (Hewitt et al., 2025).
In this paradigm, the AI doesn't just wait for a query; it maintains a mental model of the user's understanding and takes the initiative to offer unsolicited clarifications or suggest areas for exploration. This transforms the AI from a "black box" into a cooperative partner, helping humans keep pace with increasingly complex and superhuman systems.
By treating explainability as a social negotiation rather than a static broadcast, these researchers are moving toward AI systems that are not just transparent, but truly understandable.