"Explanation" is Not a Technical Term: The Problem of Ambiguity in XAI

"Explanation" means different things to different people: The paper identifies a fundamental problem in AI transparency—there is no shared definition of what an "explanation" actually is. This disconnect means AI systems might produce explanations that satisfy technical requirements but fail to help the people who actually need to understand them, such as regulators, patients, or affected communities.

Explanations must match their audience and purpose: The authors emphasize that useful explanations depend on three factors: the functional role (why the explanation is needed), the user's existing knowledge, and what information is available. A doctor needs different explanations than a software engineer, and current XAI systems often ignore these distinctions.

XAI can create false trust without establishing trustworthiness: A critical warning from the paper is that providing explanations—even detailed ones—can make people trust AI systems more without those systems actually being more trustworthy or accurate. This is particularly dangerous in high-stakes domains like healthcare or criminal justice where misplaced confidence could cause real harm.

The field lacks agreed-upon metrics for measuring explanation quality: The paper calls for XAI to develop concrete standards to evaluate whether explanations actually work for their intended purpose. Without such metrics, there's no way to systematically determine if an explanation is genuinely helpful or merely appears to be.