Assessing Collaborative Explanations of AI using Explanation Goodness Criteria

Abstract

Explainable AI represents an increasingly important category of systems that attempt to support human understanding and trust in machine intelligence and automation. Typical systems rely on algorithms to help understand underlying information about decisions and establish justified trust and reliance. Researchers have proposed using goodness criteria to measure the quality of explanations as a formative evaluation of an XAI system, but these criteria have not been systematically investigated in the literature. To explore this, we present a novel collaborative explanation system (CXAI) and propose several goodness criteria to evaluate the quality of its explanations. Results suggest that the explanations provided by this system are typically correct, informative, written in understandable ways, and focus on explanation of larger scale data patterns than are typically generated by algorithmic XAI systems. Implications for how these criteria may be applied to other XAI systems are discussed.

When Self-Driving Fails: Evaluating Social Media Posts Regarding Problems and Misconceptions about Tesla’s FSD Mode

“…Despite this, posts showed a tendency to support resolution (up to 30.6% of statements). However, many of the resolutions remained perfunctory, and specific motivations [26] may be needed to encourage resolutions with more complete reasoning traces.…”

“…Some limitations include the possible biases in reporting, the timeliness of these reports for systems that are constantly improving, the possibility that users are mistaken or incorrect about their reports or their explanations, and the unstructured and ad hoc nature of distributed social media. Some of these limitations might be avoided by developing a more formal and structured SQA system, such as the CXAI concept [26]. In such a system, additional structure might provide better verification by experts or officials, easier methods for showing agreement, and other schemes to encourage broader participation.…”

“…Many SQA platforms focus on bug fixing (such as StackOverflow or StackExchange), and they can be associated with 'how-to', 'why', and 'why-not' explanations [25]. An earlier evaluation of the observations from the CXAI system showed that a considerable proportion of posts involve describing 'What', with a smaller but substantial proportion answering 'Why' questions [26]. 'What'-style posts describe problems and limitations, but do not offer fixes or workarounds.…”

“…Previously, Mamun et al (2021aMamun et al ( , 2021b [3,26] examined how a collaborative SQA platform (CXAI) could be used to help users understand the limitations of an AI system. However, those studies used a special-purpose platform for a simple image classifier that was used only for evaluating the CXAI system.…”

“…Resolution: Whether the problem was resolved. We included this coding to determine how many of the posts resulted in satisfactory answers to queries, which Mamun et al [26] determined occurred relatively rarely in CXAI systems.…”