X-ToM: Explaining with Theory-of-Mind for Gaining Justified Human Trust

Explanations work best as a two-way conversation: Rather than simply telling users why an AI made a decision, the framework treats explanation as an iterative dialog where the AI learns what the human wants to know and adjusts its responses accordingly, making explanations more relevant and useful.

AI should model what humans understand and believe: The framework explicitly tracks three mental models—what the human intends to learn, what the human thinks the AI knows, and what the AI thinks the human understands—to generate explanations tailored to each person's actual knowledge level rather than one-size-fits-all explanations.

This approach can measure "justified" trust, not just trust: By comparing whether the human's understanding of the AI actually matches the AI's real capabilities, the framework distinguishes between blind trust (believing in AI without understanding) and justified trust (understanding why the AI works), which is critical for responsible AI deployment.

The method was tested on practical visual tasks: The researchers demonstrated the framework on real-world applications including image classification, action recognition, and pose estimation, suggesting the approach could be applied to AI systems users encounter in practice.

Theory of Mind makes explanations more natural: By drawing on psychological concepts of how humans understand other minds, the explanations feel more natural and intuitive for both expert and non-expert users to understand how complex AI models actually work.