This special track explores human trust and transparency of AI systems by investigating the concept of trust calibration in AI and related concepts and methods. Calibration of an AI model pertains to the degree to which the model accurately estimates and communicates its reliability to decision-makers, thereby informing its trustworthiness. Trust calibration, conversely, relates to the extent to which decision-makers rely on the model when it is correct and distrust it when it is wrong, thereby avoiding potential cognitive biases. While eXplainable AI has been proposed as instrumental for achieving trust calibration and avoiding potential cognitive biases, recent research highlights how it may also exacerbate these issues. While xAI strives to make AI decisions understandable to humans, this quest often leads to a paradox (White Box Paradox) where increased transparency may result in reduced trust calibration in opposing directions, both for the unveiling of AI outputs’ inherent uncertainty and complexity and for an excessively persuasive power of otherwise ‘wrong’ explanations. This track aims to attract research, including theoretical and empirical contributions, that fosters human-centred design, evaluation, and implementation of xAI support and its role in trust calibration. Discussions will focus on designing AI and xAI systems that promote appropriate (calibrated) reliance, ensuring that users are neither overly sceptical nor blindly trusting of xAI systems and their suggestions, underscoring the significance of a responsible, human-centred approach.
Topics
Examination of how xAI methods influence trust in human-AI interactions (LIME, SHAP, and Counterfactuals versus trust assessment) |
Empirical and conceptual studies of XAI methods application on trust dynamics (e.g. trust evolution time linked to user familiarity with AI systems) |
Empirical research on trust calibration in both real-world and simulated settings |
Empirical research on trust calibration (matching the user’s trust level with the actual AI-system performance) |
User feedback loops, continuous performance monitoring, and adjusting the level of automation for XAI |
Examining the xAI’s role by employing calibration information (e.g. expected calibration error, estimated calibration index, brier scores) |
Developing frameworks and metrics for measuring over-reliance (automation bias, algorithm appreciation, deference to algorithms) or under-reliance on XAI methods |
Examining the impact of explanations of AI-based systems on decision-making and trust (e.g. increased trust via higher understandability or lowered trust due to biases/errors) |
Novel approaches/techniques for XAI system design to improve the calibration of trust via AI explanations |
Investigating unexpected effects of xAI methods on trust like the White Box Paradox (i.e., increased over-reliance or under-reliance following system explanation) |
Investigations of the White-Box paradox and its effect on XAI on trust |
Interactive explanation tools, model confidence metrics, and uncertainty visualizations for avoiding automation bias in XAI |
Methods for uncertainty visualization for xAI trust calibration (e.g. takes on activation maps, vague Visualizations) |
Novel visualisation methods for representing certainty in AI predictions (e.g. confidence intervals, error bars, or heat maps) |
Visualisation for the over/underconfidence level of a model |
Ethics, accountability, and transparency assessment for calibrating trust of AI models (e.g. using Algorithm Impact Assessments (AIA) questionnaires and interviews for affected stakeholders). |
Formalisation of the broader societal implications of trust in XAI (via e.g. taxonomy of potential biases) |
Definition of societal implications of trust for XAI (e.g. via new socio-ethical impact assessment procedure elaborating on Algorithmic Impact Assessments) |
Drawing societal implications from the emergence of specific human-AI reliance patterns for XAI |