Accessibility Tools
Overly permissive use of artificial intelligence in higher education risks rewarding polished outputs over deep understanding and placing too much of a burden on students. Instead, universities must lead with pedagogical design that aligns learning, assessment and meaningful use of AI tools.
“I use AI myself, so why shouldn't my students?”
This is a familiar refrain in faculty discussions. But there is a crucial difference between a lecturer and a student. One already has the expertise that using AI productively depends upon, whereas the other is still building it.
This creates dangers, as permissive AI policies diminish opportunities for students to develop knowledge in their domain. Our concern is not with AI designed for pedagogical scaffolding, but with its use without a clear educational function.
It is therefore the responsibility of teachers, as established experts in their respective fields, to align their assessment strategies, including the permitted or prohibited uses of AI, with their pedagogical goals. This should clarify which aspects of a task cannot be delegated to AI and make the underlying rationale explicit.
Done well, this enables students to exercise informed judgement about when AI use supports or undermines their learning, in line with their own educational aspirations. Unfortunately, current practices too often prioritize convenience or cheating prevention over pedagogical alignment.
As AI increases the premium on expertise, teachers and higher education institutions should:
(a) clarify which tasks students must perform without AI to acquire the required expertise,
(b) design assessments that align with pedagogical goals and promote learning, and
(c) develop AI policies that support students' development of expertise.
There is a growing tendency – sometimes encouraged by institutions and instructors – to favour frictionless, efficiency-driven learning in ways that risk sidelining higher education’s core mission: cultivating intellectual growth and deep domain expertise.
This is often , or, in the case of AI use, to prepare students for their future work. Yet, this approach to learning may work directly against students' best interests and undermine the very relevance of higher education.
From a cognitive-science perspective, these trends reflect a common conflation: acquiring or retrieving information versus doing the deeper work of developing understanding by making predictions, encountering surprises and revising one's mental model (see Griffiths et al. 2024). This is not work that should be offloaded in the name of convenience.
It is easy to assume that because working with AI involves cognitive work, it must also involve learning. Prompting a model, reading the output, editing and re-prompting feels productive. But effort only builds learning when it is germane, i.e. directed at working through the steps that lead you to the answer, rather than getting an AI tool to produce it. This means deconstructing a problem and generating your own solution for each step before comparing with the correct answer.
Actual learning happens when you discover where your own reasoning has led you astray and you use this to update your mental model. However, as Lodge and Lobel (2026) argue, students often struggle to identify which kind of effort is germane, as they are still in the process of developing the necessary expertise. As a result, they are frequently unable to use AI in ways that effectively support and preserve their learning.
When teachers indiscriminately celebrate AI-generated improvements in essay quality, they forget their core pedagogical purpose.
As Jonathan Boymal writes: "To articulate an argument in writing is to discover what you think in the process of trying to say it. The labour of articulation is generative of thought, and ultimately of the self that does the thinking."
The same applies in learning activities such as active reading, programming, or problem solving: their pedagogical value lies less in flawless outputs than in the cognitive processes that produce understanding.
Drawing on the framework developed by Falk Lieder and colleagues (2025) on cognitive resource allocation, when assessments reward expert-level products rather than demonstrated learning, the convenient shortcut of AI becomes a rational response to the incentives built into assessment. A rethinking of assessment strategies is therefore required, not primarily, or not only, to create AI-proof tests, but to reward the learning process over outcomes, thereby fostering an environment that encourages cognitive growth.
Addressing this challenge, however, requires institutional support and appropriate training for teachers. This kind of investment is essential precisely because it makes visible a broader and recurring problem: responsibility is often demanded where expertise is still developing. This is especially apparent in university policy on AI use.
To reconcile AI use with their mission, many universities have developed ‘responsible AI policies’. However, these policies often shift the burden of tool limitations onto users by requiring them to:
(1) identify potential biases, which requires systematic testing;
(2) detect factual errors, even though outputs are optimized for rhetorical plausibility;
(3) avoid plagiarism, despite the absence of reliable source tracking; and
(4) account for climate and environmental impact, without access to clear energy-use data.
The level of epistemic vigilance these policies presume is demanding even for the best-informed experts. It is unrealistic for most learners. without the required expertise, it risks entrenching poor practices and undermining scientific judgement and integrity. Policies could treat this as a matter for teaching, where students’ judgement is developed. However, a recent study of 96 UK universities found that most of their policies are designed to monitor and punish students, despite using helpful-sounding language, and none of them show inherent trust in students (Illingworth 2026).
Therefore, responsible AI policies must start from how students actually learn, and not from what they are allowed to do. This means designing courses that reward the kind of effort that leads to learning, rather than outputs an AI could produce. It also means placing the responsibility for getting this right with the institution, not the individual student. In practice, this amounts to deciding where to restrict AI use or offer structured guidance, and to investing in teacher training and support.
Without guidance, we cannot expect students to use AI meaningfully, and policies risk optimizing for the wrong thing: efficient outputs at the expense of the work that builds understanding. This will erode the long-term mission of universities: to develop the domain expertise that allows graduates to exercise judgement in their fields.
Teaching students to use AI well is the wrong framing, because it is an instruction only an expert can follow. The real task is to build expertise first.
