Notes
Can You Explain Why You Did Not Recommend Something Else? An Experimental Study on the Effects of Contrastive Explanations and Person-Likeness on Trust in and Understanding of Algorithms
186Artificial Intelligence-based systems become increasingly ubiquitous but stay opaque to most users. For instance, when using recommender systems, people are confronted with outcomes which the system evaluated as suitable for the inquiry without an understanding how this inquiry was processed. In this contest, explainable AI approaches (XAI) aim to increase users’ understanding of and trust in these systems by presenting explanations on how an algorithm comes to a conclusion (Gunning et al., 2019; Samek et al., 2019; Shin, 2021). Two different but increasingly used explanation types are comparative and contrastive explanations. A comparative explanation describes why a decision was made, for instance because the input shared a certain number of attributes with the chosen outcome (k-nearest neighbor algorithm; e.g., Taunk et al., 2019). Contrastive explanations (also known as counterfactual explanations, e.g., Byrne, 2019; Guidotti et al., 2019; Verma et al., 2020) explain which changes in the input would have led to a different classification. Moreover, there is a trend of equipping technological systems with human-like features (e.g., voice, language, social roles) to enhance the system’s person-likeness which may also affect people’s understanding and trust (e.g., Edlund et al., 2008; van Pinxteren et al., 2020).
187Against this background, this preregistered study aims to investigate the effects and interplay of explanation type and person-likeness on understanding of and trust in an algorithm-based system and its output. In an empirical study with a 2 (contrastive vs. comparative explanation) x 2 (self-referencing vs. neutral formulation) between-subject design and a no explanation-baseline condition, 762 participants interacted with an online breakfast recipe recommender system. The breakfast guide gave either a contrastive, a comparative, or no explanation and referred to itself either with “I” or as “the system”. Effects on participants’ factual and perceived understanding and their trust in the breakfast guide’s skills and sincerity were measured.
188Human-machine interaction / algorithms / artificial intelligence / explainable AI / contrastive explanation / person-likeness / trust / understanding