Identifying critical design elements for increasing trust in computers through co-designing XAI for a mobile banking app
Abstract:
Explainable Artificial Intelligence (XAI) has established itself as a central tool for developing advanced AI systems in even high-stakes domains, as techniques of explanation have underpinned an increased trust in their deployment by experts. While there is excellent research and ongoing developments in this area, itwas felt that therewas insufficient understanding of how to apply XAI properly when the explanations were not designed for experts. To develop this understanding, particularly in the area of trust where XAI markets itself,
this project used a human-centred approach to develop substantiated design concepts and identify practical design elements. This was achieved by a participatory design process, producing XAI for select scenarios from a mobile banking app; stylised to resemble the Starling Bank mobile app, to improve believability for participants.
The studies involved demonstrated that interpretable “glass box” models which produce written, textual outputs in a concise, plain language manner, was the desirable explanation method for users. These levied existing financial literacy and natural language skills, rather than requiring the user learn how to read a more graphical explanation. The applicability of such explanations as supplementary diagrams was established, however participants recognised that these were not always desired, particularly in time-critical contexts, where the textual designs were considered preferable and sufficient. From our findings, we establish a few heuristic guidelines that can be considered for future research, particularly in XAI for banking. These findings demonstrate that there are indeed aspects where human-centred XAI approaches similar to the one pursued in this project can reveal the contextual nuance of an XAI deployment, which can provide valuable insight into what qualities are essential to the XAI used, enabling prioritisation and designs that contrasts in areas with broader XAI guidelines.
Download Alex's Thesis