AI-integrated programmatic assessment: a proposed framework for safeguarding cognitive development in AI-augmented learning environments

Generative artificial intelligence has unsettled a quiet assumption on which assessment in higher education has long relied: that the work a student submits is a reasonable proxy for what that student can do. When a fluent essay or a plausible solution can be produced with little independent effort, the inference from product to competence weakens. Institutional responses have concentrated on integrity, detection, and compliance, and less on what sustained reliance on these tools may do to learning. Emerging evidence shows that performance with AI and learning without it can come apart, and we argue that the more important risk is therefore developmental: that some forms of reliance may bypass or, over time, erode the effortful thinking through which durable understanding and professional judgement are built. We treat this as a testable possibility rather than an established general effect. Building on programmatic assessment, we propose AI-Integrated Programmatic Assessment (AIPA). The model samples three streams of evidence on the same construct: unaided cognition, AI-mediated performance, and the quality with which learners orchestrate AI. It treats orchestration as a professional competence to be taught and assessed rather than a behaviour to be policed, and it specifies explicit rules for interpreting differences between what a learner can do with and without AI, including band thresholds, recurrence and corroboration requirements, and a distinction between coherent and discrepant profiles. A confirmed discrepancy is read as a hypothesis about a current competence gap that prompts further sampling and a proportionate response; the term erosion is reserved for documented decline from a comparable earlier level. We position AIPA against adjacent approaches, including programme-level secure and open assessment, process-focused and interaction-trace assessment, and linked assessment designs, illustrate it through four worked cases, and set out a research agenda to test its claims.