Large language models (LLMs) are now being asked to turn complex machine‑learning explanations into plain‑language stories for people who must understand or challenge automated decisions. However, LLMs can repeat social bias that they learn from data, and no one has looked closely at how fair these stories are.

To fill this gap, a team has defined "fairness in XAINs" (explanatory AI narratives) and built a method to test it. They generate the same story in three ways: (1) the narrator includes the protected attribute (e.g., gender), (2) it is hidden, or (3) it is swapped to a counterfactual group. By comparing the results of five different statistical tests, they can trace where any group‑based differences come from.

The researchers applied the framework to four high‑risk datasets—German Credit, Saudi HR, Student Performance, and Law School—selecting 100 cases from each that were predicted unfavourably. For every case, five different LLMs produced a narrative, and the team checked how faithfully each story reflected the underlying SHAP explanation.

Faithfulness was measured on value, sign and rank accuracy for the three most important SHAP features. Feature coverage recorded how often these top features, protected attributes and secondary features appeared. Sentiment of the text was calculated with a standard opinion lexicon, giving a score between –1 (negative) and +1 (positive).

The study showed that, even in a controlled setting of a single prompt and temperature setting, significant differences in faithfulness, feature coverage and sentiment emerged across social groups. The analysis was limited to English narratives, binary splits of protected attributes, and a single prompt design. Moreover, the extraction of narrative content relied on a single secondary LLM, whose errors were verified on a subset of 100 stories.