Explainability Via Causal Self-Talk

AI explainability is often avoided in practice despite the XAI community developing many techniques, because most existing methods impose costs that the deep learning community is unwilling to pay—suggesting a gap between academic solutions and real-world adoption.

The paper proposes training AI systems to build causal models of themselves rather than applying external explanation methods, which the authors argue can satisfy both the need for explainability and the practical constraints of deep learning systems.

The method, called Causal Self-Talk (CST), trains agents to "communicate with themselves across time" and was demonstrated in simulated 3D environments to generate explanations that are both faithful to the agent's actual decision-making and semantically meaningful to humans.

Beyond explanations, the learned self-models enable new control interfaces, suggesting that explainability techniques could have dual benefits by also improving how humans can direct and control AI system behavior.

This approach represents a pragmatic shift in thinking about explainability—rather than treating explanation as an add-on feature, it integrates explanation-building into the core training process, potentially making it more compatible with how modern AI systems are actually developed and deployed.