Human-AI Systems in Medicine: Outskilling Versus Newskilling

AI is widely recognized as a tool that biomedical scientists, engineers and clinicians can, and should, use. However, what do we mean by a tool? I take the example of convolutional neural networks that learn latent statistical associations from images, but those associations can be used to different ends. I focus on two different uses in the field of medical diagnostics, what I call human-AI "outskilling" and human-AI "newskilling". Outskilling is a prosthetic human-AI activity to outperform human capacities (in Greek: prosthesis, adding) in tasks that experts can nevertheless perform well. I study computer-aided diagnostics (CADx) to detect polyps as an example of AI outskilling, which carries the risk of deskilling without a proven gain in meaningful outcomes. I term the second use "newskilling," a human-AI activity that brings forth something new (in Greek: poiesis) by using latent statistical associations to discover variables that human inference cannot detect. I study the example of AI deriving clinically relevant variables from retinal fundus images to derive "retinal age gaps" as an example of human-AI newskilling. There are two major conclusions based on this distinction: the design of AI uses, and the discernment of how and when to use them.