Measuring Information in Text Explanations

Text explanations need better evaluation standards: The paper addresses a critical gap in explainable AI—while text-based explanations (like rationales and natural language explanations) are increasingly used to help users understand AI decisions, there's no unified way to evaluate whether these explanations are actually good, making it difficult to compare different explanation methods fairly.

Information theory offers a common measurement framework: By applying information-theoretic concepts, researchers can measure two key qualities of explanations—"relevance" (how much the explanation relates to the input) and "informativeness" (how much it relates to the AI's decision)—providing a standardized way to assess different types of text explanations.

Different explanation types work differently: The study found that natural language explanations (written descriptions) involve a trade-off between explaining what the AI saw in the input versus what it decided, while rationale-based explanations (highlighting key text passages) don't show this same trade-off, suggesting they operate through fundamentally different mechanisms.

Standardized evaluation matters for responsible AI: Without rigorous, consistent ways to measure explanation quality, it's harder to ensure that AI systems are actually providing users with trustworthy and understandable information about their decisions—a key requirement for responsible AI deployment.