Explainable artificial intelligence (XAI): from inherent explainability to large language models

Transparency is critical for trust and adoption: The paper emphasizes that AI systems' lack of transparency creates reluctance to adopt them in mission-critical domains like healthcare and autonomous driving, highlighting that understanding why an AI makes decisions is essential for stakeholder confidence.

Multiple approaches exist for explaining AI decisions: The survey covers a spectrum of explainability methods ranging from inherently interpretable models (simpler, more transparent by design) to techniques that explain "black box" models after they're built, offering different trade-offs between accuracy and interpretability.

Large language models can improve explanations of other AI systems: A key finding is that LLMs and vision-language models can be used as tools to generate "high-level, semantically meaningful explanations" of other machine learning models' decisions, potentially making AI more understandable to non-technical users.

No single perfect solution exists: The paper identifies "strengths and weaknesses of state-of-the-art methods," suggesting that explainability remains an evolving challenge without a one-size-fits-all solution, and different applications may require different approaches.

Explainability is necessary to prevent harmful outcomes: XAI techniques enable users to "discern the basis of the decision and possibly avert undesirable behavior," positioning explainability not just as a nice-to-have feature but as a safeguard against problematic AI decisions.