Natural Example-Based Explainability: a Survey

Example-based explanations align with how humans naturally think: The paper emphasizes that humans learn and reason through examples and mental representations, making example-based XAI methods more intuitive and trustworthy than alternative approaches like saliency maps, which have dominated recent XAI research despite questions about their ability to accurately reflect how models actually work.

"Natural" examples from training data are more trustworthy than generated ones: The survey prioritizes methods using real examples directly from training data over those requiring generative processes, arguing that plausibility is essential for gaining user trust—a critical consideration for responsible AI deployment in high-stakes applications.

Multiple complementary explanation approaches exist beyond saliency maps: The paper identifies five main families of example-based methods (similar examples, counterfactual/semi-factual explanations, influential instances, prototypes, and concepts), each with distinct advantages and limitations, suggesting that responsible AI systems may benefit from combining multiple explanation types rather than relying on a single method.

Saliency maps have limitations that warrant exploring alternatives: The survey questions whether saliency maps—the dominant XAI visualization method—actually reflect models' internal decision-making processes, highlighting the importance of evaluating explanation methods' validity before deploying them in real-world systems where users rely on them to understand AI decisions.

Cognitive impact matters as much as technical accuracy: The paper emphasizes comparing methods not just on technical grounds but also on their "cognitive impact" on users, recognizing that responsible AI requires explanations that are both accurate and actually understandable to the people who need to trust and use these systems.