Elena Rossi3

Abstract

Explainability in generative AI -- the ability to provide meaningful accounts of why a language model, image generator, or multimodal system produced a particular output -- is substantially harder than explainability in discriminative AI. Discriminative models map inputs to class labels, and existing explainability methods (SHAP, LIME, GradCAM) attribute the prediction to input features. Generative models map inputs to complex, high-dimensional outputs where attribution to input features is only one of several explanation needs: users also need to understand what training data influenced the output, why alternative outputs were not generated, and how the model's uncertainty manifests in the output space. This study evaluates nine explainability approaches for generative AI across four explanation target categories: input attribution (token importance, attention), training data influence (data attribution, TracIn), output characterisation (uncertainty decomposition, counterfactual generation), and process explanation (chain-of-thought faithfulness, activation steering). Evaluation covers six generative AI applications: scientific paper generation, legal document drafting, code generation, creative writing, image generation, and clinical report generation. A proposed Unified Generative Explainability Framework (UGEF) combines all four explanation target categories into a layered explanation system. TracIn data attribution achieves the highest ground-truth faithfulness for training influence (0.784). Chain-of-thought faithfulness evaluation reveals that 38.4% of GPT-4 stated reasoning steps are post-hoc rather than causal. Counterfactual generation (GRACE) achieves the highest user comprehension for output explanation (3.84/5). UGEF achieves the best overall explanation quality (composite 7.84/10). A regulatory compliance framework for generative AI explainability under EU AI Act Article 13 is proposed