Researchers unveil XstrAI: AI that tailors explanations to any audience

Researchers from Italy have developed XstrAI, a framework designed to bridge the communication gap in explainable AI (XAI) by tailoring explanations to specific audiences. Traditional methods, such as SHAP (SHapley Additive exPlanations), provide numerical insights but often fail to communicate effectively across diverse groups like patients, clinicians, and data scientists. XstrAI addresses this by treating local explanations as immutable evidence and structuring their communication through a multi-agent system.

The system employs three specialized LLM agents: an audience-aware planner, a linguistic realizer, and a validator. The planner ensures the explanation aligns with the audience's expertise, the realizer generates the narrative, and the validator checks for grounding, consistency, and communicative risks. XstrAI was evaluated against 11 baselines on diabetes and stroke risk prediction tasks, using both intra-narrative fidelity metrics (alignment with SHAP evidence) and extra-narrative assessments (audience appropriateness via LLM judges and surveys). Results show XstrAI's narratives were consistently assigned to the correct audience and preferred by clinicians and patients, with competitive performance for data scientists.