After a three-decade hiatus, interest in explainable artificial intelligence has resurfaced. The concept is sometimes abbreviated as XAI (eXplainable artificial intelligence), and it has been mentioned in grant applications and the popular press. This comeback is fueled by evidence that many AI apps are underutilised or never used at all, owing to ethical concerns and a lack of confidence on the part of their users. Users will be more equipped to comprehend and trust intelligent agents if they construct more transparent, interpretable, or explainable systems, according to the running hypothesis.
While there are many ways to improve intelligent agent trust and transparency, many trusted autonomous systems will use two complementary approaches: generating decisions in which one of the criteria taken into account during the computation is how well a human could understand the decisions in the given context, which is often referred to as interpretability or explainability; and explicit decision-making.
Justification and Interpretability
We’ll go over the differences between interpretability, explainability, justification, and explanation as they’re employed in this article and in artificial intelligence.
Lipton offers a taxonomy of desiderata as well as interpretable AI approaches. The claim made by Lipton that explanation is post-hoc interpretability is adopted in this study. The degree to which an observer may grasp the cause of a choice, according to Biran and Cotton’s concept of interpretability of a model. Explanation is thus one manner via which an observer can gain knowledge, but there are clearly other modes, such as making judgments that are intrinsically easier to grasp or introspection. I use the terms “interpretability” and “explainability” interchangeably. A justification explains why a decision is sound, but it does not always try to describe the decision-making process itself.
How Do People Explain Their Behavior Using Social Attribution?
Perception is crucial to social attribution. While the reasons of behaviour can be stated at a neurophysical level, and possibly even lower, social attribution is concerned with how people attribute or explain the behaviour of others, rather than the actual reasons of human behaviour. According to Heider, social attribution is defined as a person’s viewpoint.
As a result, intentional behaviour is always contrasted with unintentional behaviour, citing that state laws, sports rules, and other regulations all treat intentional actions differently than unintentional actions because intentional rule breaking is punished more severely than unintentional rule breaking. While intentionality is an objective truth, it is also a social construct in which individuals ascribe intents to one another, whether objective or not, and utilise these intentions to socially engage.
Intelligence in the Crowd
For those working on collective intelligence, such as multi-agent planning, computational social choice, or argumentation, study on group behaviour attribution is crucial. Although this line of research appears to be less well-developed than attributions of individual behaviour, findings from Kass and Leake, Susskind et al., and, in particular, O’Laughlin and Malle that people assign intentions and beliefs to jointly-acting groups and reasons to aggregate groups suggest that the large body of work on attribution of individual behaviour could serve as a solid foundation.
Source & Reference : Paper with Code