Explainable Generative AI (GenXAI): A Survey, Conceptualization, and Research Agenda

Generative AI requires new approaches to explainability: Traditional explainability methods designed for AI systems that recognize patterns are insufficient for generative AI systems that create new content. The paper emphasizes that as GenAI applications become "increasingly more complex and multi-faceted," novel explainability needs have emerged that weren't addressed by previous XAI research.

Explanations must meet multiple new standards beyond accuracy: Beyond simply explaining why a system made a decision, the paper identifies emerging requirements for GenAI explanations including verifiability (can the explanation be checked?), interactivity (can users engage with it?), security (is it protected from manipulation?), and cost-effectiveness—suggesting that responsible AI requires explanations that are robust across multiple dimensions.

Explainability can be built in at multiple stages: The paper identifies "different avenues to ensure XAI, from training data to prompting," indicating that organizations have multiple opportunities to make generative AI more transparent—whether by carefully selecting training data, designing how users interact with the system through prompts, or other methods throughout the development process.

This is an urgent, evolving challenge: The paper's identification of "more than ten directions for future investigation" and its focus on "emerging desiderata" signals that explainable generative AI remains an unsolved problem requiring ongoing research and attention from multiple disciplines including social scientists and information systems researchers, not just technologists.