Explanation paper argues LLM outputs need context-aware justifications

A new arXiv paper revisits what makes an explanation good in the context of AI systems. The authors propose a definition inspired by counterfactual explanations while also accounting for the user’s prior beliefs.

That matters because explainability is often treated as a generic feature, but users need different explanations depending on what they already know and what decision they are trying to make. LLM outputs are especially hard to explain because their reasoning is not a simple causal trace.

The paper adds philosophical rigor to a practical adoption problem: trustworthy AI systems need explanations that are useful to the person receiving them.