Currently reading “the WONKHE report” by Dickinson & Marshall (2026)

I’ve seen lots of summaries and references to the “WONKHE report” on AI, so I had to check it out. It is based on focus group interviews and a survey with responses from more than 1k UK students, and it is a bit difficult to read because it is not always clear what results are vs the authors’ interpretations/claims, but it has very clearly structured “main findings”, so check it out yourself! Below are my personal take-aways.

The first interesting point is how students describe their own AI use. There are six main ways, and students move between the different usages depending on context:

• As search engine, to save time but also because it seems to be easier to get exactly what they want than using other ways of literature search. Which is probably true, but on the other hand it is of course not possible to describe the method of a literature search in a way that is reproducible, and it also does not contribute to training conducting of structured searches.
• To structure students’ thoughts (and this is the most commonly mentioned use). The report states “Multiple students described developing their argument independently, then asking AI to propose a format – structure as a distinct skill from argument, one that AI can assist with without touching the intellectual content” but from the AI output I have seen, I think that usually there will be some kind of “touching the intellectual content”, even though that might not be immediately obvious. Students are quoted saying that they, for example, give a handful of points and ask AI to suggest the order to present them in. Assuming that the order matters for the argument they are making or story they are telling (otherwise why wouldn’t they just go top to bottom on wherever they took their notes), how is that not “touching the intellectual content”?
• For debugging code. Here, the authors write that “The learning happens in the correction. This student was thinking in real time, not outsourcing.” While that might be the case depending on what exactly students are doing, I think oftentimes it is, in fact, outsourcing.
• As a tutor, to make it ask questions for practice purposes, to get feedback on work, as a “mock assessor” for formative feedback.
• To fill gaps in prior knowledge or when teachers don’t provide good explanations.
• To speed up production (and this is what students think universities are worried about)

The good news: contexts that influence how students use AI are mostly things that teachers can influence, possibly with the exception of time poverty (assuming that the demands we place on students’ time are reasonable already, otherwise that’s the thing to fix).

Other factors that can influence whether students use AI or not include

• whether students care about a subject or not. This can be because of their own interest (which we can, at least to some degree, spark and maintain with how we frame topics, by allowing them agency and letting them experience competence and relatedness in the classroom, …) and because of how relevant a course is to a program (which we can either motivate better, or fix if it really isn’t relevant).
• whether students think the tutor cares about them or not. We can definitely influence this! (See also Corbin et al. (2026) on recognition).
• whether students feel pressure to use AI to keep up with competition, get good grades, … I hate this framing of “if we don’t use it, we fall behind”, whether the idea is that the falling behind is behind other students or the demands of the job market…
• how assessment is being done — I am continuing this point outside of the bullet point list!

Recommendations for assessment

So, assessment. When students think that what is valued and assessed in a course is the final product and not the learning itself, they are more likely to use AI. “When marking feels unfair, AI stops feeling like a choice and starts feeling like a necessity.”

Looking into how students perceive assessment formats, the authors describe the “essay paradox“: >90% of students in the survey say that they are confident that essays can show real learning, but they also know that it is really easy to cheat in essays. “That is the design problem – an assessment format that rewards understanding when it’s present but cannot detect when it’s absent.” That becomes more clear in the focus group data, where students say that essays can be about understanding but also about bullshit bingo, i.e. putting in things that sound nice without actually understanding what they mean.

In group work, students really dislike that individual contributions become invisible and everybody ends up with the same grade for the product, and it stresses them that they don’t know how their peers work with AI and that they might be held responsible for something they didn’t do and weren’t aware of.

The authors summarize how students describe properties of real learning: “abstract knowledge meeting concrete application with real stakes, the presence of other people whose engagement makes going through the motions insufficient, personal stake in the material, and memorability” and good assessment then has the characteristics of “visibility of individual understanding, application, accountability, relevance, and developmental feedback“. Assessment formats that do work to show learning, according to students, are portfolio-type assignments with reflections that lead to documented iterative development of a challenging task, or ongoing conversations, working with applications of learning to relevant scenarios, … The authors write that “[t]he design challenge is to build accountability moments that are experienced as developmental conversations rather than surveillance events – and that are accessible to students for whom live performance under pressure is itself a barrier.”

One way to make sure that students understand that it is about learning and not just about a product is to let them know that they will have to talk about the product, the process, and their learning, and to show their understanding. Students say that they would often not be able to explain work that they have submitted as their own if they were asked to. If they know that they will have to explain, though, they use AI in a way to prepare them for that. If not, they often use AI to “optimize” their submission and thus grade. And students do say that they want accountability moments, but they also say they are afraid of them. The recommendation in this report is to introduce live verification moments, but to make sure that students are prepared for them through formative attempts and low-stakes practice and to make sure that formats are accessible or that there are alternative formats available for those who need them.

In summary for assessment design, the authors recommend asking ourselves the question “how would a student prove to us that they understood this, rather than simply produced something plausible?”

AI policies & educational injustice of AI declarations

AI policies are present in most places, but usually not detailed enough to give students useful information about what they are actually allowed to do or not. Nevertheless, in a “choose all that apply” question, 75% of the students tick “the assignment guidelines” when making decisions about using AI, 72% what feels right ethically (and in the examples, it becomes clear that very different criteria are used here, for example if it can help someone else), 66% what would help them learn, only 26% consider the risk of detection.

But: Students report often very different guidelines and advice from different teachers in different courses, which is confusing for students, and one of the recommendations is to coordinate AI guidance within programs (but, me thinks, students need to also understand the complexity and context-dependence of making decisions around AI and that there is no one right answer, but it depends on context and ethical considerations and so much more, so maybe they just have to live with different guidances in different courses (assuming, of course, that those guidances are clear!)). And even if the guidelines are the same across courses, they are still hard to interpret. The meaning of phrases such as “your own work” is not clear. Is it that it’s the students original ideas, or writing, or final product but not process, …? “Policy that uses [the “your own work” framing] as its operative standard is not providing guidance. It is outsourcing the ethical decision to the individual student while maintaining plausible institutional deniability.” This is one reason for why AI declarations do not work — it is just impossible to exactly define all possible use cases. Another is that they are “creating a perceived penalty for honesty rather than an incentive for transparency“. The authors write “The combined effect is damaging – students who use AI legitimately either lie on declarations or avoid AI entirely to avoid the perceived risk of declaring. Students who use AI most heavily are presumably the least likely to declare honestly. The declarations are catching the wrong people.” They elaborate that “The de facto norm is being set by the most risk-tolerant students and the most permissive lecturers. The policy burden falls entirely on the conscientious.” Unclear policies have an emotional toll on learners, but so does the (perceived) requirement to use AI. “For some students, it is closer to moral injury – a sense that using AI, even legitimately, compromises something about who they are as learners.”

What they also found which I think we don’t consider enough is that “AI policy is affecting recruitment.” We don’t know how big the effect of AI policy on decisions to attend a specific university or course actually is, but similarly to students seeking out LU because of its placement in sustainability rankings, some students might actually make decisions based on whether AI use is encouraged and supported!

On the topic of policies and institutional work, I like the recommendation to “[a]udit what AI is compensating for” (for example insufficient library systems), and to use that information to fix it, and the one to “[t]reat workload, curriculum relevance, and equitable access to AI tools as AI policy questions“. For example, there is a >20% gender gap in reported usage in this report, and quite often based on a “principled decision“. How should we deal with that? The report recomments to “Create structured opportunities for students to develop and share their ethical thinking around AI. The most thoughtful students in this research have built personal doctrines about AI use that are more sophisticated than most institutional policies. That intellectual work is currently invisible and unsupported. Seminars, peer discussions, or case study workshops where students surface and compare their positions would do more than another round of policy revision – and would treat students as partners in working this out rather than subjects to be governed.” This is also important because “The social dynamics of AI use are polarised. With a substantial cohort of students vocally opposed to any AI use and another substantial group actively using it for planning, studying, and writing support, there’s a real social divide. Students who use AI may feel reluctant to discuss it openly given the strength of anti-AI sentiment among their peers. This could push AI use underground – the opposite of the transparency that institutions say they want.” So more talking about ethics with students!

Also, there is the issue of AI as disability-related cognitive support, where “AI is filling a support gap, not a learning one“. Students report that AI helps with retention, so they are making “a learning outcome claim rather than a process claim“. Surely then policies need to account for that?

What do students really want to learn?

Lastly, one very interesting point. 61% of students who say they can pass without understanding also say that AI made them question what they are learning, compared to only 24% of students who don’t believe that. What is it that students wished their studies helped them develop? “The pattern is not “give us harder academic challenges.” It is something closer to – help us become people who can do things in the world. The desire for confidence in one’s own judgement, named by 29 per cent, sits directly against a system that elsewhere in the survey students describe as primarily rewarding conformity to tutor preferences and rubric compliance.“ (my emphasis).

Dickinson, J., & Marshall, M. (2026). Trained to stop learning. How students are experiencing assessment and learning in an age of AI. WONKHE Report. https://wonkhe.com/wp-content/wonkhe-uploads/2026/03/Trained-to-stop-learning-F.pdf

A little Sunday wave watching from just before the summer vacations. Fascinating how quickly waves are dampened by algae in the water near the surface!

And now add some floating barriers that dampen out waves and protect shelter from the wind!

Funny how big an influence such a small floating rope can have!

Love the smell of those roses!