ExplainabilityTrustworthinessTransparencycs.CV
Towards a Novel Measure of User Trust in XAI Systems
Miquel Miró-Nicolau, Gabriel Moyà-Alcover, Antoni Jaume-i-Capó, Manuel González-Hidalgo, Adel Ghazel, Maria Gemma Sempere Campello, Juan Antonio Palmer Sancho
May 9, 2024
Key Takeaways
-
XAI methods aim to build user trust by making AI decisions transparent: The paper addresses a critical gap where deep learning models make decisions without explaining their reasoning, and explainable AI (XAI) is being developed specifically to help end-users understand and trust these automated systems.
-
A new metric combines both performance and trust factors: Rather than measuring only whether an AI system works accurately, the proposed trust measure incorporates both how well the system performs and explicit trust indicators, recognizing that user confidence requires more than just correct answers.
-
The methodology shows improved sensitivity across different scenarios: The three case studies demonstrated that this novel trust measure is more responsive to varying situations than existing approaches, suggesting it can better capture how trust changes depending on context and use case.
-
Trust measurement is essential for refining XAI systems: By establishing a concrete way to measure trust in explainable AI, organizations can systematically evaluate and improve their XAI implementations rather than relying on subjective assessments of whether explanations are helpful.
Abstract
The increasing reliance on Deep Learning models, combined with their inherent lack of transparency, has spurred the development of a novel field of study known as eXplainable AI (XAI) methods. These methods seek to enhance the trust of end-users in automated systems by providing insights into the rationale behind their decisions. This paper presents a novel trust measure in XAI systems, allowing their refinement. Our proposed metric combines both performance metrics and trust indicators from an objective perspective. To validate this novel methodology, we conducted three case studies showing an improvement respect the state-of-the-art, with an increased sensitiviy to different scenarios.