Explanation in artificial intelligence: Insights from the social sciences

In the citing paper

“There can be multiple ways of making data and code accessible, and there can be multiple ways to justify a decision and different degrees of transparency in the decision (Miller 2019; Mittelstadt et al. 2019).”

In the citing paper

“For instance, machine-learning algorithms can be programmed in a way they can explain based on what data and which decision rules certain predictions were reached, this is called explainable AI (xAI).”

In the citing paper

“Also, user experience (UX) in XAI, specifically Human-AI collaborative decision-making, may require more attention than the explanation ease of use [47].”

In the citing paper

“Some users prefer a detailed explanation, while others need a brief explanation based on their curiosity level [47].”

In the citing paper

“To the best of our knowledge, there is little to no empirical research on the effectiveness of these two humancentered explanations derived from the literature (i.e., Adadi and Berrada, 2018) using Miller's criteria in fostering end-users' trust in AI.”

In the citing paper

“Counterfactuals combine Miller's selectivity and contrastivity criteria.”

In the citing paper

“we identified feature importance explanations and counterfactual explanations as two promising post-hoc explanations for achieving HCXAI.”

In the citing paper

“Social science literature suggests that natural human explanations are predominantly qualitative [63]-while people may have an intuition about which feature is relevant to the decision, or which feature is more relevant than another, they may not be able to articulate a precise quantification.”

In the citing paper

“Tim Miller argues for the situation of AI explanations within social and human sciences literature on cognition and decision making, given that the problems XAI addresses are fundamentally those of the relationships between people rather than people and technology [ 43 ].”

In the citing paper

“HCI skills such as philosophy, psychology, and cognitive science have proven their worth in the field of XAI by developing explanations that were able to stimulate the human explanation process [89].”

In the citing paper

“In addition, they state that explanations are selective , meaning that focusing solely on the main causes of a decision-making process is sufficient.”

In the citing paper

“Explanations that explicitly contrast correct diagnoses against plausible alternatives (contrastive explanations) may be particularly valuable for teaching differential diagnosis skills [159].”

In the citing paper

“This finding challenges the common belief that counterfactual explanations, which simulate AI reasoning, are easily understandable and enhance trust [24].”

In the citing paper

“Interventions that cue users to reflect on their goals and strategies via question prompts or conversational interfaces have been shown to improve outcomes in the workplace [92,122,209], education [51], and other domains [17,147].”

In the citing paper

“Moreover, further complications are introduced when AI systems produce outputs in interaction with humans or other AI or complex sociotechnical systems (Borrego-Díaz & Galán-Páez, 2022;Miller, 2019).”

In the citing paper

“Most of the empirical research on algorithmic transparency has focused on the how, reflected in the explainability, interpretability and, ultimately, in the observability of algorithms (Miller, 2019;Robert et al., 2020;Shin, 2021;Shin & Park, 2019).”

In the citing paper

“xAI refers to the explainability of specific decisions, but can also refer to how a model works and develops (e.g., Miller 2019).”

In the citing paper

“Such explanations thus satisfy the selectivity criterion proposed by Miller (2019) because they show how certain factors influenced a decision.”

In the citing paper

“However, studies show that many XAI methods are designed according to developers' intuition of what constitutes a 'good' explanation, rather than actual insights into users' needs [73].”

In the citing paper

“Much of the literature still speaks in subjective terms of 'good' or 'quality' explanations [73], often without measuring how well those explanations integrate into clinicians' decision-making processes.”

In the citing paper

“The human trait of explaining the underlying reasoning behind a decision and applying this same logic and requirement in the context of AI is complex (Miller 2019).”

In the citing paper

“First, explanations are better when constrictive , meaning that a prerequisite for a good explanation is that it does not only indicate why the model made a decision X, but also why it made decision X rather than decision Y.”

In the citing paper

“It is also explained that probabilities are not as important as causal links in order to provide a satisfying explanation.”

In the citing paper

“The concept of what defines a good explanation is not trivial and is an aspect that is widely studied in the literature (Miller, 2019; Holzinger et al., 2020).”

In the citing paper

“In an important contribution, Miller (2019) emphasises the importance of respecting explanatory practices as people experience them-in particular that good explanations are contrastive, i.e. that”

In the citing paper

“social functions have to do with building or maintaining interpersonal relationships, creating shared meaning (Miller, 2019), and with supporting actions or societal practices (e.g.excusing or blaming others, or responding to decisions that affect one) (Mittelstadt et al., 2019;Wachter et al., 2018).”

In the citing paper

“The rapid growth of AI has accelerated discoveries across diverse scientific fields and affected every work environment [62–65] and is reshaping the labor market with unprecedented speed and scale, with 40% of the global workforce expected to require AI, enhancing the need for significant AI upskilling or reskilling [66].”

In the citing paper

“Also, creating AI with the ability to explain and justify its decision-making process [5,38,44] could increase the end-users willingness to accept the AI's recommendations [53,55].”

In the citing paper

“In three studies [67], [128], [137], the classifiers used in the experiments are not specified.”

In the citing paper

“Contfactual explanations, as a mixture of tabular and visual output representations, appear also in an augmented reality framework [137].”

In the citing paper

“However, the two aforementioned studies [128], [137] lack an evaluation of the quality of the generated contfactual explanations themselves.”

In the citing paper

“Miller argues that a good explanation is contrastive, selective, and truthful.”

In the citing paper

“However, some research asserts that this approach is the most intuitive for human users to comprehend [7].”

In the citing paper

“because humans tend to devalue explanations that contradict their prior beliefs.”

In the citing paper

“which are not only intuitively understandable for humans [7]”

In the citing paper

“Generative AI can provide personalized learning plans and tools based on students' individual needs and learning styles to improve learning efficiency and effectiveness.”

In the citing paper

“However, this personalized learning can also lead to unfair outcomes, as different students may face different opportunities and challenges.”

In the citing paper

“In addition, if generative AI models only focus on students' individual needs and interests, and ignore some important learning objectives and standards, this may also lead to students' knowledge and skills being less comprehensive and in-depth (Miller, 2017).”