HCC: Medium: Improving Human-AI Collaboration on Decision-Making Tasks

Harvard University, Cambridge MA

Investigators

Abstract

From loan approval to disease diagnosis, there are many situations in which human decisions are being assisted by artificial intelligence (AI). For example, a clinical decision support system might suggest a possible diagnosis or highlight a potential medication interaction based on elements of the patient's history that the human doctor might have otherwise missed. It was expected that by combining the complementary strengths of people and AI systems (human+AI), the quality of the decisions made in such settings would be better than that of either people or machines alone. Unfortunately human+AI systems have not lived up to this promise: Even with explainable AI, human+AI systems often perform worse than either alone. Recent work shows that users of AI decision-support often have a superficial understanding of the AI. This leads to inappropriate levels of trust swinging from ignoring the AI to over-reliance. This project will create human+AI systems that perform better than either alone. The research team will develop and test specific tools and techniques that will be valuable for creating effective human+AI decision systems across many domains. The project will explore three ways of improving AI-based decision support. Humans typically engage AI systems heuristically, while successful interaction calls for an analytical approach by the human partner. Only then can the human appropriately combine their knowledge with the AI recommendation and its explanation. To encourage more analytic engagement, the project will design and test (a) adaptive cognitive forcing functions: cognitive interventions that guide the human to pay closer attention the AI's information (applied only when most valuable to avoid frustrating the user), and (b) intelligent contrasts: methods that ground the AI's information as a contrast to what the human is likely to do. The latter will spark the human user's curiosity about why the AI may be recommending something different than the human. The last thrust involves building systems to help users understand the AI in the context of the data that power it, enabling a more global understanding of when the AI is likely to be useful. This project will explore specific versions of each approach described above applied to clinical treatment decision and to nutrition planning. The research results will enhance our understanding of how to create better human+AI teams. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.