Files
Download Full Text (85.3 MB)
Publisher
Technological University Dublin
Description
In 1950 when Alan Turing first published his groundbreaking paper, computing machinery and intelligence and asked “Can machines think?” a new era of research exploring the intelligence of digital computers and their ability to deceive and/or imitate a human was ignited. From these first explorations of AI to modern day artificial intelligence and machine learning systems many advances, breakthroughs and improved algorithms have been developed usually advancing at an exponential pace. This has resulted in the pervasive use of AI systems in the processing of data. Concerns have been expressed related to biased decisions by AI systems around the processing of personal data in domains such as recruitment, medicine and the judicial system, AI systems which make life changing decisions for users. However, legislation such as the EU’s AI act has called for greater regulation of artificial intelligence in the EU including catagorising according to risk, developing systems with greater transparency and the right to an explanation of AI systems’ decisions. Explainable artificial intelligence (XAI) can be best described as a model which produces details or reasons which make the functioning of an AI system clear or easy to understand. Much research and development has been done in this area to demystify the black box nature of some AI systems. With the right to an explanation XAI will play a leading role in the compliance of high-risk AI systems for companies and in the delivery of explanations for those who engage with AI.
However, when an XAI method is presented to a user how do we know that the user understands the explanation? When we factor in other metrics which might be crucial when evaluating users’ understanding of an XAI output such as; user satisfaction, user curiosity or need for an explanation, user trust in the system following an explanation and users’ mental models of the AI system we can see the multiple evaluation methods, scales and tests which may need to be considered. Also, bear in mind the breath of user types who require explanations and the volume of domains which utilise AI in automated decision-making which may also influence the evaluation method employed.
Evaluation methods traditionally used within the IT industry for software and websites have been utilised, examined, evaluated and verified extensively. Although there are many standardised scales and evaluation methods used to evaluate software and websites such as SUS, PSSUQ and SUMI, few of these methods translate specifically to the domain of XAI. Those designed specifically for XAI such as SCS, goodness check, satisfaction scale, and trust scale have not been as thoroughly tested and validated as those for software and websites. Also, many qualitative evaluation methods such as AAR/AI, explainability scenarios and counterfactual scenarios specifically tailored for evaluating AI and XAI suffer the same fate. This is to be expected since XAI in comparison to software and websites is a relatively new field of research.
As part of a larger study, we present an overview of quantative methods in the form of standardised scales and qualitative techniques considered user experience methods which evaluate more traditional forms of information technology such as software or websites but with an emphasis on those which are best suited and have been validated in the evaluation of XAI. We also present an overview of evaluation methods specifically designed to evaluate XAI and discuss how these might be used in conjunction with traditional evaluation methods to determine users understanding of XAI outputs.
Publication Date
2023
Keywords
Artificial Intelligence, Explainable Artificial Intelligence (XAI), mixed methods approach
Disciplines
Artificial Intelligence and Robotics | Computer Sciences
Conference
2023 Irish Human Computer Interaction (iHCI) Symposium
DOI
https://doi.org/10.21427/T327-AJ03
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 4.0 International License.
Recommended Citation
Sheridan, H., O'Sullivan, D., & Murphy, E. (2023). Unlocking the Black Box: Evaluating XAI Through a Mixed Methods Approach Combining Quantitative Standardised Scales and Qualitative Techniques. Technological University Dublin. DOI: 10.21427/T327-AJ03