CX-ToM: Counterfactual Explanations with Theory-of-Mind for Enhancing Human Trust in Image Recognition Models

Counterfactual explanations are more effective than attention maps: The research demonstrates that traditional heat map-based explanations used in most AI systems are insufficient for building human trust. Instead, the CX-ToM framework uses "fault-lines"—explanations showing what specific features would need to be added or removed from an image to change the AI's decision—which better help humans understand model reasoning.

AI explanations work best as conversations, not one-time answers: Rather than providing a single explanation, CX-ToM frames explanation as an iterative dialogue between human and machine. This back-and-forth communication process allows the system to adapt explanations based on what the human understands, similar to how humans naturally explain things to each other.

Understanding both human and machine "minds" improves explanations: The framework uses Theory of Mind to model what the human user intends to understand, what the machine actually "thinks," and what the human believes the machine thinks. By explicitly accounting for these different perspectives, the system can bridge gaps in understanding and provide more targeted explanations.

Semantic-level features matter more than pixel-level details: CX-ToM focuses on meaningful, human-understandable concepts (like "stripes on a zebra" or "pointed ears on a dog") rather than low-level visual patterns. This approach makes explanations more interpretable and actionable for non-technical users trying to understand why an image recognition model made a particular decision.