Source-linked AI summary
Not All Explanations Are Sought: Information-Seeking Psychology for Human-Centered XAI
Andrea Beretta, Salvatore Rinzivillo
TL;DR
HCXAI lacks a unifying account of why users seek, avoid, or scrutinize explanations, a gap that is especially consequential for agentic systems. The paper applies information-seeking psychology, proposing that instrumental, hedonic, and cognitive utilities are distorted by cognitive biases. It concludes that XAI should design for explanation motivation—when and why users want to know—rather than treating explanation availability or transparency as sufficient.
Problem
HCXAI has underexplored why and when people seek explanations, despite evidence that users vary in engagement and may over-rely on or avoid explanations.
Method
The paper applies Sharot and Sunstein’s information-seeking framework to HCXAI, analyzing instrumental, hedonic, and cognitive utility and biases that distort their evaluation.
Results
The framework identifies excessive information-seeking that fragments attention and insufficient information-seeking that leaves critical risks and misunderstandings unexamined.
Takeaways & Limitations
HCXAI should shift from making explanations available to designing for when and why users actually want to know.
Abstract
from arXiv · showhide
This position paper argues that human-centered explainable AI (HCXAI) should incorporate insights from the psychology of information seeking. Drawing on Sharot and Sunstein's framework of information-seeking motives, we propose that people evaluate whether to engage with explanations based on three types of expected utility: instrumental (will it help me act better?), hedonic (will it make me feel better?), and cognitive (will it improve my understanding?). Each utility is estimated through a lens shaped by well-documented cognitive biases, including illusion of control, automation bias, unrealistic optimism, impact bias, overconfidence, and confirmation bias. These biases can lead to two failure modes: excessive information-seeking that fragments attention without improving decisions, and insufficient information-seeking that leaves critical risks and misunderstandings unexamined. This challenge is particularly acute for agentic AI systems, where explanations must support not just understanding a single output but anticipating cascading actions, assessing risks, and deciding when to intervene. By integrating information-seeking psychology into HCXAI, we advocate for a shift from making explanations available to making them sought: designing systems that account for when and why users actually want to know.
1 INTRODUCTION
HCXAI has advanced explanation design but has underexplored why and when people seek explanations. The paper frames information-seeking psychology as necessary for understanding varied engagement, overreliance, and explanation avoidance.
- HCXAI research has underexplored why and when people actually seek explanations from AI systems.
- XAI often assumes users want explanations, although explanation value and user engagement vary across individuals and contexts.Recent findings include overreliance on readable LLM explanations and compliance under time pressure regardless of explanation quality.
- Making explanations available or accessible does not ensure that users will seek or scrutinize them.
- People selectively seek or avoid information according to its expected effects on actions, emotions, and understanding.The paper treats this selectivity as a fundamental feature of cognition rather than a failure of attention or literacy.
- This gap is especially important for agentic AI, where explanations must support anticipating cascading actions, assessing risks, and deciding whether to intervene.Sophisticated explanations still fail to support oversight when users avoid, ignore, or dismiss them.
- The paper proposes integrating information-seeking psychology into HCXAI to explain engagement patterns and design systems around when and why users want to know.Its framework uses instrumental, hedonic, and cognitive utility while accounting for biases that distort those evaluations.
2 HOW PEOPLE DECIDE WHAT THEY WANT TO KNOW
The paper applies an information-seeking framework in which people evaluate explanations through instrumental, hedonic, and cognitive utility. Cognitive biases can distort each evaluation and shape whether information is sought, avoided, or treated as redundant.
- Information-seeking is driven by instrumental, hedonic, and cognitive utilities.Instrumental utility concerns action and decisions, hedonic utility concerns anticipated emotion, and cognitive utility concerns strengthening a mental model.
- Instrumental utility reflects whether information is expected to improve decisions or enable more effective action.People avoid information they perceive as useless for action.
- Hedonic utility reflects the anticipated emotional impact of knowing, encouraging good-news seeking and bad-news avoidance.
- The illusion of control and automation bias distort instrumental utility by overestimating intervention or over-relying on automated systems.Automation bias can reduce the perceived value of understanding why a system acts.
- Unrealistic optimism and impact bias distort hedonic utility by reducing motivation to seek information about risks, failures, or uncertainties.
- Illusion of knowledge and confirmation bias distort cognitive utility by encouraging redundancy judgments or selective confirmation of existing beliefs.
3 IMPLICATIONS FOR EXPLAINABLE AI
A psychology-of-information-seeking perspective challenges the assumption that more transparency is inherently better. It identifies excessive and insufficient information-seeking as failure modes with direct consequences for decisions, oversight, and human-AI teaming.
- Current XAI often equates providing information with providing knowledge and prioritizes completeness and transparency.This view overlooks that information is not always perceived as useful or sought.
- People decide what to know based on expected utility rather than passively engaging with all available information.
- Applying the framework reveals excessive and insufficient information-seeking as two XAI failure modes.Excessive seeking fragments attention without improving decisions, while insufficient seeking leaves critical risks and misunderstandings unexamined.
- The illusion of control can prompt users to seek explanations believing that understanding will enable intervention even when intervention is impossible.Explanations may increase perceived understanding, which users mistakenly equate with capacity to intervene.
- Avoiding explanations about uncomfortable risks, failures, or uncertainties can create blind spots where scrutiny is most needed.
- Excessive information can generate frustration and fragment the user’s mental model, producing worse outcomes despite more explanation.The paper attributes this to failing to consider whether and when the user needs the information.
- Effective human-AI teaming requires explanations that account for perceived instrumental, hedonic, and cognitive utility.Otherwise, users may disengage into blind over-reliance or seek excessive clarification that disrupts task completion.
4 FUTURE DIRECTIONS
The paper reframes HCXAI around users’ motivation to seek explanations and identifies future research on how utility trade-offs shape explanation timing, engagement, and design.
- HCXAI should move beyond treating explanations as inherently valuable by examining how perceived utility and cognitive biases affect reliance and human-AI teaming.
- The paper shifts HCXAI’s unit of analysis from explanation content to explanation motivation and presents this shift as a foundational theoretical lens.
- For agentic AI, future work should study whether explanations are best offered before, during, or after multi-step execution, because each timing supports different goals and risks.
- High-stakes settings create tension between instrumental needs for actionable explanations and hedonic avoidance of distressing risks, so optimizing one utility may compromise another.
- Individual differences in how users weight the three utilities could inform adaptive explanation strategies for agentic AI and future HCXAI systems.