Generative and agentic AI have changed not just how students complete assessments, but also what those assessments can credibly evidence. Cohen Ambrose, Course Director at the Digital Learning Institute in Ireland, argues that higher education needs a learning-theoretic framing for this moment.

We cannot credibly assert that students understand a new field or body of knowledge upon successful programme completion unless we have collected verifiable evidence that they can operationalise that knowledge and skillset in more than one novel context. This was true before generative and agentic AI. It is true with more force now. 

Whenever we measure what a student can do, the assessment context and its constraints shape what we see. Capability is observed through a context, not behind one. What we call “understanding” is always inseparable from the constraint regime we witness the student navigate. A nursing student who passes a medication calculation exam in a quiet hall has demonstrated something. What they have demonstrated is not that they can calculate safely on a ward at three in the morning during a cardiac arrest. If the only context we ever give students to perform competency in is the same as the one in which we taught them, we are not measuring understanding. We are measuring rehearsal. 

This is not an academic integrity or security problem. It is a learning-theoretic problem about what our assessments are entitled to claim. 

A learning theory for the AI moment 

Yan and Gašević’s (2026) recent proposal for a new learning theory they call Agentivism describes what AI is doing in human cognition. Human agency is now distributed across a sociotechnical ecology. A graphic design student who briefs Midjourney, rejects three iterations, intervenes in the fourth and composes the final piece, has not produced an artefact that maps cleanly onto their unaided cognitive labour. Neither has the engineering student who debugs with Claude, nor the law student who pressure-tests their argument against an LLM before submission. The question shifts from “did the student use AI?” to whether traceable human agency persists through their work. 

This dissolves the unit of analysis (the learning artefact) on which most current assessment practice rests. If the artefact is a joint production of student and AI, then asking whether the student “really” produced it is the wrong question. The right question is whether something cognitively coherent on the student’s side of the ecology can be traced through what was produced. 

Beyond the integrity and security problem 

Neither of the two frames institutions currently rely on can answer that question. 

Here in Ireland, the National Academic Integrity Network’s own guidelines diagnose the systemic causes of misconduct as overlapping deadlines, perceived loss of learning value, lack of opportunities for resubmission and absence of consequence (NAIN 2023, pp.18-19). Every cause identified is a property of the system. Yet the same document asks educators to “emphasise values such as integrity, trust, and truthfulness” (NAIN 2023, p.9) and asks students to live those values themselves. The institution diagnoses its own design failures and assigns the remediation work to the students harmed by them. This is not values education; it is an implicit liability shift. 

Fawns et al. (2024, p.4) distinguish two strands institutional discourse tends to conflate: academic integrity as a values-based mission, and assessment security as the engineering of rules that enforce certain forms of engagement. Both try to answer the same question, whether the evidence we hold warrants the claim we make about a student’s competence, and both answer it in the wrong vocabulary. Integrity asks whether the student behaved honestly. Security asks whether the conditions prevented dishonesty. Neither asks the question Agentivism forces on us: does traceable human agency cohere across what the student submitted? 

What needs to take its place 

To answer that question, we need to define what coherence actually is. 

Juarrero (2023) describes coherence as the signature of constraint closure. By constraints, she means the commitments, frames and dispositions that shape how a system engages with the world: what it attends to, what it ignores, how it weights competing considerations, when it adapts and when it holds firm. A learner’s constraint structure is the architecture of those commitments. It is the set of things a business student treats as relevant when analysing a strategic decision, the way they balance cost against ethical exposure, the questions they ask before trusting a piece of data. Coherence is the condition by which those constraints hold together across situations that demand adaptation rather than reproduction. In practice, we observe this condition in degrees. A second-year student’s constraints may cohere across two contexts and fracture in a third. The work of assessment is to read where that fracture happens, and what it tells us. 

Coherence becomes visible only under contextual variation, as I have argued in my newsletter Architecting Uncertainty. A snapshot cannot reveal it. Novel-context transfer is the only condition under which coherence manifests at all. 

A learning-theoretic frame, grounded in this definition, does not abandon assessment security; it absorbs it. If the evidence of a student’s learning does not cohere, they do not earn the degree. The mechanism remains intact, but the vocabulary changes. The institution still decides whether the evidence warrants the credential. What changes is the ground on which that decision rests. Where integrity asked whether the student behaved honestly, and security asked whether the conditions prevented dishonesty, coherence asks a different question: given the evidence in front of us, can we defensibly conclude that this student has formed the durable capability the credential is meant to certify? 

The security apparatus most institutions rely on (invigilation, prohibition, detection, surveillance) cannot do this work because each of these instruments operates within a single context. They can secure a snapshot. They cannot read coherence, because coherence is not a property of any single moment. 

This reframes what the university certifies. Not the artefact a student submits, but the constraint structure their cognition has formed, evidenced by whether that structure survives perturbation across contexts the assessment regime deliberately introduces. That a student created the artefact with AI does not invalidate it as evidence of their competencies, so long as the constraint structure their cognition has formed (including in its collaborations with AI) verifiably survives perturbation across novel contexts. Certification reads competence. Competence is inferred from evidence. If the evidence does not cohere, the inference fails, and no credential follows. 

A question for university leaders 

For too long, university leadership has placed the burden of verifying durable student capability on academic and support staff alone. In the era of Agentivism, we require a broader strategic and systemic approach; one that takes seriously the change underway in the global knowledge and skills economy. This work goes beyond the purview of an academic integrity unit, a programme, or a set of national guidelines. It is brave, complex, uncomfortable work. Do not put it on the shoulders of individual staff. Resource it properly, redesign whole systems to allow it to scale, and lead with it. At the current moment, you cannot credibly claim that a student holding a degree parchment with your seal and signature can do what that piece of paper guarantees. What will you do about that? 

Cohen Ambrose is Course Director and Learning Experience Designer at the Digital Learning Institute in Ireland. His work focuses on assessment design for the era of generative and agentic AI, with a particular interest in novel-context transfer, constraint-based learning theory, and the architecture of uncertainty in higher education. He writes about learning transfer and other related topics in his Substack newsletter, Architecting Uncertainty. 

References 

Fawns, T., Bearman, M., Dawson, P., Nieminen, J.H., Ashford-Rowe, K., Willey, K., Jensen, L.X., Damşa, C. and Press, N. (2024) ‘Authentic assessment: from panacea to criticality’, Assessment & Evaluation in Higher Education, pp. 1-13. doi:10.1080/02602938.2024.2404634. 

Juarrero, A. (2023) Context changes everything: how constraints create coherence. Cambridge, MA: MIT Press. 

National Academic Integrity Network (2023) Generative artificial intelligence: guidelines for educators. Dublin: Quality and Qualifications Ireland. Available at: https://www.qqi.ie/sites/default/files/2023-09/nain-generative-ai-guidelines-for-educators_1.pdf (Accessed: 16 May 2026). 

Yan, L. and Gašević, D. (2026) ‘Agentivism: a learning theory for the age of generative and agentic AI’, arXiv preprint, arXiv:2604.07813v1.