# Explainable AI without Interpretable Model

**Authors:** Kary Främling  
**arXiv:** [2009.13996](https://arxiv.org/abs/2009.13996)  
**Date:** 2020-09-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2009.13996v1/

## Key points

- Explainability is a challenge in the field of Artificial Intelligence (AI)
- AI systems need to provide explanations for their reasoning, especially when end-users are affected by their decisions
- Complexity of AI models, trained using machine learning techniques, makes explainability difficult
- Existing methods for Explainable AI (XAI) focus on extracting interpretable models, but they may not accurately represent the original black-box model and may not be understandable for end-users
- The concept of Contextual Importance and Utility (CIU) enables the production of human-like explanations directly from black-box outcomes without relying on an interpretable model
- CIU explanations accurately map to the black-box model itself and can be applied to any type of black-box system without being constrained by specific models or algorithms
- CIU incorporates the utility concept from Decision Theory, providing a new dimension to explanations compared to traditional XAI methods
- CIU allows for producing explanations at different levels of abstraction and using various vocabularies and means of interaction, tailored according to specific contexts and target users' needs
- The CIU framework provides accurate mappings between black-box outcomes and human-like explanations while incorporating the utility concept for a more comprehensive understanding
- This advancement improves transparency and trust in AI systems, particularly in critical decision-making processes that impact end-users.

## Summary

Explainability has always been a challenge in the field of Artificial Intelligence (AI). However, with the increasing use of AI systems in society, it has become more crucial than ever for these systems to be able to provide explanations for their reasoning. This is particularly important in situations where end-users are affected by the decisions made by AI systems, such as being rejected from a job application or having a loan application denied. One of the main difficulties in achieving explainability lies in the complexity of AI models, especially those trained using machine learning techniques. These models often contain numerous parameters that are difficult to analyze and comprehend, leading to their characterization as "black-box" systems. Many existing methods for Explainable AI (XAI) focus on extracting an interpretable model that can generate explanations. However, these interpretable models may not accurately represent the original black-box model, and their understandability for end-users remains questionable. To address these challenges, this paper introduces the concept of Contextual Importance and Utility (CIU), which enables the production of human-like explanations directly from black-box outcomes without relying on an interpretable model. CIU explanations accurately map to the black-box model itself and can be applied to any type of black-box system without being constrained by specific models or algorithms. In addition to feature importance, CIU incorporates the utility concept from Decision Theory, providing a new dimension to explanations compared to traditional XAI methods. This allows for a more comprehensive understanding of how and why certain decisions were made by AI systems. Furthermore, CIU offers flexibility in producing explanations at different levels of abstraction and using various vocabularies and means of interaction. This adaptability ensures that explanations can be tailored according to specific contexts and target users' needs. Overall, this paper presents an innovative approach to achieving explainability in AI systems without relying on interpretable models. The CIU framework provides accurate mappings between black-box outcomes and human-like explanations while incorporating the utility concept for a more comprehensive understanding. This advancement has significant implications for improving transparency and trust in AI systems, particularly in critical decision-making processes that impact end-users.