Black-box AI in medicine: a standard of care without interpretability

Thesis

Artificial intelligence (AI) systems, especially deep learning models consisting of many hidden layers, have gained widespread attention over the last decade for their seemingly sudden jump in performance. This attention has been followed by widespread experimentation with such models, and ultimately mainstream application, across almost every domain of human effort. They have been used across the sciences, across industry, in medicine, law, finance, and even across the humanities and arts. Their uses have further ranged from being a decision-support systems, an aid to human operators, and autonomous recommender systems across retail, social and traditional media. Over the past decade, there is scarcely an area of human activity that has not seen their uptake, and sometimes even dominance. Medicine has been one domain where their performance, especially in evaluating medical imaging, has seen substantial improvement in the past decade. This thesis will examine the extent to which the use of such deep learning models in medicine, due to their black-box nature, poses novel ethical and epistemic problems. It will ultimately argue that such problems are either not present, or can be overcome – contrary to the more widespread discussion in the literature (and in policymaking). This introduction will lay the groundwork for this analysis by identifying and discussing some key concepts and themes, and articulating the broad argumentative strategy of the thesis.

Authors: Mishra, A

• DOI: 10.5287/ora-kergzd5k6
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: medicine; artificial intelligence; ethics
• Subjects: Artificial intelligence; Ethics; Philosophy