JMIR Publications

ABSTRACT

Artificial intelligence (AI) assistants are moving from single-encounter advice toward continuing support for people who manage a chronic disease. Most evaluations ask whether people decide better while the assistant is present. Far fewer ask what people can do later without it, or whether that later capability should change how support is given. This Viewpoint argues that these are two separate evidence questions. The first concerns an assistance footprint: the average causal difference in later unaided capability between alternative support histories, for a specified task, population, and follow-up period. Randomized experiments in education and in general adult samples show that the way AI is used can change later unaided performance. Available evidence has not yet established whether alternative histories of AI support produce clinically meaningful differences in later unaided self-management capability, or whether anticipating such differences improves subsequent support decisions. The second question is whether a decision policy that anticipates a footprint makes better support choices than an otherwise identical policy that does not, and than a flexible policy that learns from each person's full history. Using peripheral artery disease as a worked example, the paper shows why similar effects can justify different decisions for different people, and why independence from AI is not an automatic goal. It outlines a staged, clinically governed evaluation, states what each possible result would and would not mean, discusses counterarguments, and offers a minimum reporting list for studies of repeated AI support. The contribution is not a new theory, whose components are established, but a precise causal question, a demanding comparator, and a practical evaluation agenda. Until such evidence exists, assistance footprints are a research construct and not grounds for reducing care.