Explainable AI Reloaded: Challenging the XAI Status Quo in the Era of Large Language Models

The traditional "black-box opening" approach to AI explainability is becoming obsolete with Large Language Models. The paper argues that the original XAI vision—making AI systems transparent by understanding their inner workings—is increasingly impossible for non-experts, requiring a fundamental rethinking of how we approach AI transparency.

A human-centered approach to explainability is more practical than an algorithm-centered one. Rather than focusing solely on technical explanations of how LLMs work internally, the paper advocates shifting focus to what users actually need to understand and trust these systems, acknowledging the "epistemic blind spots" of purely technical explanations.

Explainability can happen outside and around the black-box, not just inside it. The paper identifies three alternative dimensions for achieving explainability: examining factors external to the algorithm, focusing on system boundaries and interfaces, and leveraging the broader infrastructure surrounding AI systems—all of which may be more achievable than opening the black-box itself.

We need to reset expectations about what explainability means in the LLM era. The paper challenges stakeholders to accept that complete algorithmic transparency may not be feasible or necessary, and instead develop realistic frameworks for responsible AI that work within these technical constraints.