When Stability meets Sufficiency: Informative Explanations that do not Overwhelm

Explanations need to balance detail with simplicity: The research identifies a key tension in AI explanations—while minimal explanations are easier to understand, they often leave out too much information for users to actually evaluate whether the AI system is working correctly. The proposed method addresses this by providing a sequence of explanations at different levels of detail rather than just one.

Stability matters as much as accuracy: The paper emphasizes that good explanations should be "stable," meaning small changes to an input shouldn't drastically change the explanation. This is critical for building trust, since users need to know that explanations reflect genuine patterns in how the AI works, not just random artifacts.

One-size-fits-all explanations don't work across different types of data: The research validates their approach across three different data types—images, tables, and text—suggesting that responsible AI systems need explanation methods flexible enough to handle diverse real-world applications, not just one specific domain.

Users need to trace the AI's decision-making process: Rather than showing a single explanation, the proposed method (PSEM) traces a path from the full input down to the minimal information needed for a decision. This "journey" helps users develop better intuition about how the model actually behaves locally, which is essential for meaningful human oversight.

Comprehensibility requires both fidelity and sufficiency: The research confirms that effective explanations must be both faithful to how the model actually works AND sufficient to justify its decisions—neither criterion alone is enough for users to properly understand and evaluate AI systems.