Using LLMs for Explaining Sets of Counterfactual Examples to Final Users

Counterfactuals can help people understand how to change AI decisions: The paper demonstrates that by showing "what-if" scenarios where specific factors are changed, AI systems can explain not just why a decision was made, but how someone could potentially change that outcome—making explanations more actionable for real people facing automated decisions.

Large language models can translate complex data into plain language: The research proposes using LLMs to convert sets of counterfactual examples into natural language explanations, addressing a key challenge: raw counterfactual data is difficult for non-technical users to interpret, but narrative explanations may be more understandable.

The approach uses step-by-step reasoning to improve explanation quality: Rather than asking an LLM to generate explanations directly, the pipeline breaks the task into smaller steps that "mimic human reasoning," suggesting that structured prompting can produce more coherent and accurate explanations than single-step approaches.

Evaluation of AI explanations remains incomplete: While the authors developed a "closed-loop evaluation" method to assess whether explanations accurately reflect the underlying counterfactuals, they acknowledge that "further experiments with other datasets and human evaluations should be carried out"—indicating that real-world validation with actual users is still needed before deployment.

This work bridges causality and usability in AI accountability: By combining causal inference methods with natural language generation, the research addresses a gap between technically sound explanations and explanations that are genuinely useful to people affected by automated decisions.