Björn Fasterling (EDHEC): “If AI is to improve the learning experience, it must first be stopped from undermining it”

Your main concern regarding AI is the potential loss of learning experience. Could you tell us more about that?

Björn Fasterling: At EDHEC we regularly have lively discussions among academic colleagues, students and the school’s management on how to best make use of AI technologies to enhance the students’ learning experience. It is necessary to have this discussion. The promises regarding “scalable personalised learning experience” are tempting and yet, right in the word “experience” is where I see a major issue.

While AI expands what we can produce, it does nothing to expand our capacity to attend. Human attention remains biologically limited, no matter how much AI increases our output. Something has to give. My concern is that what gives is the experience itself. If attention is a limited capacity that helps determine what we process and experience, then being able to “do more” with the same amount of attention may mean experiencing less of what we do. And if we experience less, we also have less experience to share with others. 

Learning, however, depends on actively engaging with one’s work, reflecting on one’s achievements, and participating in an environment that allows exchange with teachers and peers. From that perspective, AI seems less to open new possibilities for learning than to threaten the conditions that make learning possible.

According to you, while being aware of its counterproductive effects, doesn’t AI enable many possibilities for students and teachers?

Björn Fasterling: Indeed, there is growing evidence that AI can have counterproductive effects on learning (1), which means that a professor like me must be cautious when integrating it into teaching, feedback, grading and the broader classroom experience.

At a very abstract level, the solution to “scaling personalised learning” without degrading the learning experience may seem straightforward. 

AI, one might say, should function as a “cognitive enabler” rather than as a tool for “cognitive offloading”. It should enrich classroom experience through AI-powered simulations rather than replace supposedly boring lectures. It should be used to generate problems that students then solve without it, rather than to outsource problem-solving itself. It should strengthen attention rather than erode it, increase motivation rather than encourage disengagement, foster critical reflection rather than invite the unreflective adoption of AI-generated output, stimulate creativity rather than stifle it, and support memory retention rather than weaken it. This list could easily be extended. 

However, this “should-rather-than” framing is misleading. It suggests that AI has no inherent tendency in one direction or the other, and that everything depends on how well we choose to use the “tool”. The evidence so far points to a more difficult reality. 

Although AI can support learning under certain conditions, it also carries a real tendency towards forms of use that undermine engagement, attention and critical reflection. Steering it in the right direction therefore requires deliberate pedagogy, as well as financial, technical, and organisational resources, and sustained effort from both teachers and students. All of which are in short supply.

Can you give some examples of the difficulties and effort needed? For example, why have so many professors abandoned take-home assignments, and what has been lost in the process?

Björn Fasterling: Almost every academic colleague I speak with has encountered student assignments that appear to be written largely with the help of large language model (LLM) tools such as ChatGPT, Gemini, Claude or other applications. The problem is not really that students may be cheating, but rather that they may be suffering a loss of experience when they outsource cognitive work. They may spend less time writing text, but they also lose the learning that comes from struggling with argumentation, structure, and language.

When students really think through an assignment problem, write their own solutions, and submit them to me, they share their learning experiences with me. I can assess them, respond to them, and offer feedback that connects to what the students have understood. If the assignment is instead a collage of prompted text, there is much less experience being shared, and much less that I can meet with relevant feedback.

Many teachers, myself included, have therefore stopped relying on assignments as a method of assessment and have turned instead to alternatives such as oral exams, in-class written exams, continuous assessment, and feedback through individual discussions with students. Yet assignments had a particular value. They enabled students to demonstrate a wide range of skills and competencies that are more difficult to assess through these alternatives.

What about providing feedback? Couldn’t AI realistically automate the kind of qualitative, personalised feedback for students in your courses on ethics or law, for example?

Björn Fasterling: If you have a high number of students in your class and teach a subject such as business ethics or human rights law, where assessment often turns on students’ ability to reflect critically on a problem, analyse it, and propose possible solutions, providing personalised feedback quickly becomes tedious and very time-intensive. As a result, too often, teachers cannot provide detailed individual feedback that students deserve. This is where the prospect of automating qualitative feedback appears as a way of making it both “personal”, “fair” and “scalable”, while at the same time relieving teachers from weeks of mind-numbing grading.   
The idea sounds simple, and it could even work in a limited sense. With the help of AI, it is possible to automatically generate a grade and a justification for that grade. One could feed an LLM-based AI with course materials and a detailed discussion of an exam problem, while carefully indicating the aspects on which students are expected to be assessed and the weight assigned to each of them. However…

However?

Björn Fasterling: Well, such automated feedback is neither personal nor fair. It is not personal because AI essentially replaces human feedback. In a sense, the teacher’s feedback is somewhat embedded in the instructions given to the AI. Yet that does not change the fact that the process interrupts the communication and shared learning experience between teacher and student. In that respect, automated qualitative feedback resembles the more traditional automated feedback associated with multiple choice tests. As a grading supplement, that is not necessarily problematic. But as a basis for justifying an overall grade in a course, it is, sorry to say, quite impersonal.

Then, such qualitative automated feedback is difficult to regard as really fair, first, because LLM-based systems do not always produce stable evaluations. As a result, the same piece of work may plausibly receive different assessments, and even different grades. In an educational setting, such variability is problematic, since fairness presupposes a sufficient degree of consistency in the treatment of comparable cases. Of course, variation can, and too often does, occur with human evaluators.

However, I would advance the hypothesis that the judgment of an experienced teacher who has sufficient time to appreciate student work is likely to be more consistent than that of an AI system. In any case, consistency and variation is perhaps not even the main point in regard of fairness. A potentially bigger problem is that even with the best instructions and prompt design many subtle but decisive aspects of student writing and performance may go unrecognised. For example, from time to time, I have seen students come up with ideas that I have not considered in beforehand, yet which prove convincing and sometimes even reveal extraordinary ability. An AI system would likely penalise the student’s deviation from the preset grading instructions, whereas I, as a professor, would reward them for their originality.

Moreover, I think that automated feedback merely addresses a problem that should not arise in the first place. The issue may not be that providing feedback is too tedious or time-consuming for professors. It may instead be that classes are simply too large, and that personalised, fair, and qualitative feedback cannot easily be scaled.

We would like to go back to the positive effects of AI. Doesn’t it, by carrying out tedious data intensive tasks, free up time for really human activities?

Björn Fasterling: I have often heard and reheard the argument that AI would “merely” automate tedious, data-intensive cognitive tasks. From this perspective, automation would “free” humans for more context-sensitive, critical and strategic work and would also increase the importance of distinctly human qualities such as empathy, love, and social connection. 

A historic parallel of technology diminishing the relevance of one cognitive capacity but “unleashing” many others is the invention of writing. Writing reduced the need to memorise, yet it advanced civilisation by enabling storage and transmission of information, so that human cognitive effort could be shifted to other activities. 

The analogy to writing is interesting but has limits. Writing arguably made one cognitive experience (memory retention) less relevant but expanded many others. Contemporary AI tools, by contrast, risk displacing an increasing range of cognitive experiences at once: writing, summarising, translating, drafting, and more, to the point that there are not many types of cognitive experience left that could be expanded. Moreover, the claim that AI will lead humans to cultivate empathy and love does not appear very compelling to me, at least not for now. It is not obvious that gains in automation are matched by gains in empathy and love. In some contexts, the trend clearly runs in the opposite direction, as illustrated by the prospect of cheaper warfare enabled by AI-powered drones.

Ok, but at least, doesn’t AI allow more space for critical thinking? Couldn‘t this be leveraged in the classroom?

Björn Fasterling: In an educational setting, one might suppose that students can outsource tedious memorisation, drafting, and data collection, allowing them to focus more fully on qualitative tasks, including critical thinking. I would say, however, that it may be difficult to reach the stage at which one can engage in critical thinking if one has become accustomed to outsourcing unappealing cognitive tasks to AI. 

Yet even if that were not the case, I still find the “unleashing logic” difficult to accept. Gains in efficiency of output are not necessarily, and perhaps only rarely, matched by greater willingness to think critically about a problem, analyse possible courses of action, and reflect on their consequences. Since students and professors, too, tend to optimise their time, critical thinking may be sacrificed in order to “free up” more time for what appears more relevant, urgent or enjoyable. 
In that sense, AI does not “unleash” critical thinking. Rather, the increased use of AI requires teachers to make a proactive effort to safeguard critical thinking, which may indeed to some extent include integrating AI into the classroom. For example, I sometimes ask students to engage in adversarial dialogues with AI, or I ask AI to generate a problem that students must solve without its assistance, or to draft a proposal for a legal norm that students must then assess (2). All this may be well and good, but does not really expand, let alone “unleash”, critical thinking in the learning environment. I could perfectly do interesting critical thinking exercises without AI.

Thank you, Björn. We should continue this discussion at a later time! But for now, what is your conclusion? 

Björn Fasterling: Yes, I agree, the discussion needs to be continued. With new insights coming in at such a rapid pace, it’s not unlikely that I will change my mind on some of the points we talked about. For now, I would say, if AI is to improve the learning experience, it must first be stopped from undermining it. Countering its negative effects takes time, effort, and resources. That is why the promise of efficiency gains in education should be met with caution. AI may not produce real efficiency gains in the learning experience so much as just shift required effort elsewhere.

References

(1) There are numerous studies. Here are some recent ones: 

Kosmyna, N. et al. (2025), “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task”, at: https://doi.org/10.48550/arXiv.2506.08872

Moșoi, A.A. et al. (2025). “Do students need to think hard? The interplay of AI and cognitive abilities in solving problems.” Educ Inf Technol 30, 24337–24364 (2025). https://doi.org/10.1007/s10639-025-13738-8

Martínez-Gordon, A. et al. (2026). “Evaluando la deuda cognitiva en la educación asistida con IA.” Advances in Building Education, 9(3), 33-43. https://doi.org/10.20868/abe.2025.3.5649

Krsmanovié, Gala, and Fadi P. Deek. (2025).  "Self-reported cognitive effects of AI applications on learning." Journal of Higher Education Theory and Practice 25 (1): 55-62. https://articlegateway.com/journals/index.php/JHETP/article/view/7563

For an overview over further studies: Jose B. et al. (2025) The cognitive paradox of AI in education: between enhancement and erosion. Front. Psychol. 16:1550621. https://doi.org/10.3389/fpsyg.2025.1550621 

(2) Some use cases and problems regarding AI and critical thinking in management education are discussed by Larson, B.Z. et al. (2024). “Critical Thinking in the Age of Generative AI”. Academy of Management Learning & Education 23 (3), https://doi.org/10.5465/amle.2024.0338