Where and when: Thursday, June 17 at 2-3pm in 303S-561
Speaker: Mike Merry, PhD student, supervised by Pat Riddle, and Jim Warren
Abstract: Wide-ranging concerns exist regarding the use of black-box artificial intelligence (AI) in sensitive contexts. Despite performance gains and substantial hype, uptake of AI is still hindered by these concerns. The ability to explain the inner workings of a black-box model is one way to help alleviate these concerns and much work has been done in the space. However, sometimes inconsistent definitions developed from a computer-science perspective have not necessarily addressed the on-the-ground aspects of these concerns.
We present a mental-models inspired definition of explainability for artificial intelligence with a focus on healthcare. Teamwork in human contexts provides a basis for understanding how we should interact with AI in complex scenarios. Mental models provide a strong basis for analysing and understanding that teamwork, but this is not how explainability is currently defined. Here, we provide a new definition of explainability that aligns with mental models, especially defining the context and putting it at the centre of the evaluation of explanations.