Core Concepts
Generative AI (GenAI) has shown remarkable capabilities, but its widespread adoption raises the need for explainability to enhance transparency, accountability, and user control. This work identifies key challenges and desiderata for Explainable GenAI (GenXAI), and proposes a taxonomy to categorize existing and future XAI techniques for GenAI.
Abstract
The article discusses the importance of Explainable Artificial Intelligence (XAI) for Generative AI (GenAI) and the key challenges in achieving effective GenXAI. It highlights several novel aspects that make XAI more crucial and challenging for GenAI compared to pre-GenAI AI systems.
The core motivation for GenXAI includes the need for users to adjust and verify GenAI outputs, the increased reach and impact of GenAI across diverse user groups, the difficulty in evaluating GenAI systems automatically, and concerns around security, safety, and accountability.
The key challenges for GenXAI stem from the lack of access to model internals, the interactive and complex nature of GenAI systems, the difficulty in evaluating explanations, and the risk of ethical violations in explanations.
The article then proposes several novel and emerging desiderata for GenXAI explanations, including verifiability, lineage, interactivity, personalization, dynamic explanations, consideration of costs, alignment with criteria like helpfulness and harmlessness, and ensuring security and conveying uncertainty.
Finally, the article presents a taxonomy for categorizing GenXAI techniques based on the explanation properties (scope, modality, dynamics), as well as the input and internal properties of the XAI methods (foundational source, access requirements, self-explainers, sample difficulty, and dimensions from pre-GenAI XAI).
Stats
"GenAI has the potential to unlock trillions of dollars annually."
"ChatGPT was the fastest product to reach 100 million users and still rapidly grows."
Quotes
"Explainable AI has also been identified as a key requirement to support prompt engineering by experts."
"LLMs are known for many shortcomings, ranging from generating harmful, toxic responses to misleading, incorrect content."
"Providing explanations should not compromise security, as insights on GenAI models might facilitate attacks or exploit vulnerabilities."