What do we actually measure when students use AI?

There are two common ways of dealing with AI without actually dealing with it. One is to allow it without changing anything else. The other is to ban it. The ban is understandable as an impulse, but it doesn't answer the question of how the course should be designed to make learning visible. It simply kicks the design problem down the road.

A study by Castro-Lopez et al. (2026) shows that students who use AI on their own initiative perform worse—not because AI is inherently harmful, but because its use lacks a pedagogical framework. The course is simply not designed for it. Hausman et al. (2025) describe a parallel pattern in a policy analysis from the European research network CEPR: grades rise when AI becomes common in courses, but the skills that those grades traditionally signaled do not. The signaling value and the actual knowledge are drifting apart. Meanwhile, Ranganathan and Ye (2026) note that AI does not free up time; it intensifies work. What was an acceptable final result yesterday is merely a starting point today. This logic applies to our students just as much as to ourselves. When we allow, or assume, that students use AI, the conditions for the entire course change. But if the learning outcomes, the activities, and the assessment remain the same as they were before AI existed, we are no longer measuring what we think we are measuring. We are measuring something else, and we don’t quite know what it is.

In this context, constructive alignment  becomes a necessary tool, not just a pedagogical buzzword. John Biggs' idea is simple: learning outcomes, learning activities, and assessment must be interconnected. The student should practice what is being assessed, and the assessment should measure what the outcomes point toward. When this chain holds, the course is coherent and fair. When it breaks, gaps appear that the student rationally fills with whatever is at hand—and today, AI is what is at hand.

It is within these gaps that two troublesome phenomena arise. 'Shadow learning' is built up when the course does not acknowledge the existence of AI, or fails to incorporate a pedagogically sound way of working with it. The student uses it anyway, but outside the teacher's field of vision and without the framework that could turn that usage into actual learning. 'Facade learning' is something else. It is the ability to produce something that looks like knowledge but is merely the handling of digital tools. Well-formatted, but substantively empty. Sejdiu and Sejdiu (2025) note that 86 percent of students' AI use in assignments went undetected by teachers, and even experienced assessors struggle to distinguish AI-generated text from a student's own. Shadow learning is a design problem. Facade learning is an assessment problem. Both are best managed through a course where the chain of alignment is well thought out.

The quickest solution when the AI issue arises is to reintroduce on-campus hall examinations. A controlled environment, no aids, problem solved. It is an understandable impulse, and in some contexts, it is a perfectly reasonable form of assessment. However, when a hall exam is chosen as a response to AI rather than as a response to the learning outcomes, it becomes a control measure for identity verification—a way to ensure that the correct student is writing, but nothing that ensures the examination measures what the course's learning outcomes intend. Not all learning outcomes can be measured over a few hours in a quiet room. The ability to reason, evaluate sources, or communicate to different audiences requires formats that hall exams rarely allow.

Ultimately, course development in the light of AI is also a matter of fairness. When we raise the requirements for what students are expected to produce, we assume that everyone encounters AI tools on roughly the same terms, which is rarely the case. A student with established AI habits produces more with the same effort, while one who lacks that experience faces the same elevated demands without having been given the opportunity to develop the competence required to meet them. Higher education is one of the few arenas that can counteract this inequality, but that requires us to take responsibility for training students in AI use—not as just another course module, but as part of how we think about what it means to be proficient in a subject today.

This is not a task that can wait until next semester. What is the first thing you need to reconsider in your own course?

Text written by Lars Johansson, Head of Division, Akademus

Sources

Castro-Lopez, A., El Abed, M., Cervero, A., & Álvarez-Blanco, L. (2026). From AI adoption to underperformance? Investigating the impact of interactive AI tools on student outcomes in higher education. European Journal of Higher Education. https://doi.org/10.1080/21568235.2026.2620686

Hausman, N., et al. (2025). Generative AI in universities: Grades up, signals down, skills in flux. CEPR/VoxEU.

Ranganathan, A., & Ye, X. M. (2026, February). AI doesn't reduce work, it intensifies it. Harvard Business Review.

Sejdiu, N. P., & Sejdiu, S. (2025). The quiet transformation of higher education in the AI era. Open Research Europe.