Not All Explanations Are Sought: Information-Seeking Psychology for Human-Centered XAI

People don't automatically want explanations, even when they're available. The paper argues that users evaluate whether to engage with AI explanations based on three types of expected value: whether it will help them act better (instrumental), feel better (hedonic), or understand more (cognitive). This means simply providing explanations isn't enough—systems must be designed to make people actually want to seek them out.

Cognitive biases can lead to two opposite problems: people either ignore explanations they should read or obsessively seek information without improving their decisions. The paper identifies specific biases like overconfidence, automation bias, and confirmation bias that cause users to either skip critical explanations or get stuck in excessive information-seeking that fragments their attention. For high-stakes AI systems, both extremes are dangerous.

AI explanations are especially critical for systems that take multiple actions over time, not just single decisions. The paper emphasizes that agentic AI systems (those that act autonomously) require explanations that help users anticipate cascading effects, assess risks, and know when to intervene—a much higher bar than explaining a one-time prediction.

Responsible AI design should shift from "making explanations available" to "making explanations sought." Rather than assuming users will read explanations when needed, designers should understand the psychology of when and why people actually want information, then build systems that encourage appropriate information-seeking behavior at critical moments.