Fabrice Cavarretta, professor in the Management Department and in charge of the PhD seminar on Pedagogy at ESSEC, co-author of the book Prière de rendre votre écosystème moins absurde (Payot, 2024).

Generative AI, particularly its text-generating capabilities, has gained significant attention in recent years. Experts often identify the education sector, specifically pedagogy, as an industry that may face significant changes due to AI.

The Potential Impact of AI on Classrooms and Evaluations

After decades of evolution, computers have become a standard object of pedagogical environments, whether in lectures or examinations. For sure, the lingering question regarding the loss of attention that screens create in class situations remains. Yet, computers often appear at some point or another of participants’ experiences, including in evaluative processes, whether for preparing an at-home report or conducting some in-class examinations.

However, many pedagogues resent that generative AI might create a particular problem in such evaluative processes. Current Large Language Models (LLMs), typically ChatGPT or Claude, have the capability to produce content that may constitute a valid response in an evaluative process, typically by being able to produce essays responding to questions in most disciplines. GenAI can successfully mimic human answers at university level.

If AI can provide apparently valid answers, it threatens the integrity of pedagogy, specifically in its evaluative parts. This triggers an apparently legitimate request: Why don’t we forbid AI in those parts of the process?

Is this feasible?

This institutional question arises mostly because, individually, pedagogues are at a loss to enforce such bans. They have become used to asking for typed examinations, and assumed that essay writing triggered limited cheating: in class, communication across students was assumed limited; at home, getting external help was assumed limited. Those assumptions could be challenged, yet are not the subject of our current discussion.

With generative AI, educators realize that the ease and speed with which it provides answers is just too tempting. Another assumption that this discussion leaves aside is whether honor codes could address this problem, since they are supposed to prevent misbehavior and have demonstrated their efficiency in many universities. Hence, institutions that rely on honor codes might succeed at harnessing AI usage by just asking participants to commit to not using it in designated sequences, typically during evaluations. However, many institutions do not rely on honor codes.

Computers are now well entrenched in our processes, difficult to remove them. Hence arises another fantasy by which the pedagogical institution could magically remove AI from computers. However, this is increasingly unlikely as those algorithms are more and more provided natively in most computing environments. Even the throttling of internet access would not ensure absence of genAI.

The worry is clear and widespread: generative AI will likely generate cheating, and no normative nor technical mechanism might attenuate it sufficiently.

The True Problem: Preparing for Tomorrow’s Work

However, this line of questioning misses the most important societal issue: education should prepare participants for future activities, whether they be scientific, societal, or industrial.

The emergence of AI is therefore not only a challenge to the evaluation of students, but it might also radically change the nature of the very activities for which we train them. Sciences do change with AI, as much as humanities do change with AI, and definitively doing business does change with AI. Therefore, the interesting question is how should pedagogy change, given the change in the downstream activity?

For sure, no one knows today what the exact nature of future work will be, given the impact of AI. But this has always been the case for pedagogy. We train students not knowing exactly what task they may precisely have to accomplish 2 to 15 years down the road. Yet, pedagogy has so far successfully functioned through a proper mix of first-order activities (counting, planning, etc.) and second-order activities (critical thinking, culture, etc.), a mix supposed to prepare students for the future. How to evolve this mix considering the emergence of AI ?

Aviation model

To address this question, let’s consider an industry that encountered similar upheavals with the introduction of computers, but only a few decades earlier: aviation.

Historically, for most of the 20th century, planes were flown exclusively with the direct input of humans. Humans flew the plane: they pushed buttons and touched the stick to get the plane to do what the human cognitive system was deciding in minute details. One thought of the pilot would correspond to a conscious mechanical reaction in the aircraft. Computers were there in the cockpit, but mostly providing information such as visualization, trajectory estimation, etc.

Then a few decades ago computers entered “the loop”, i.e., the computer started mediating the relationship between the pilot and the plane. Nowadays, many military planes require the constant supervision of computers to remain stable. Computers control the link between human inputs and machine reaction. And this has spread to many civilian planes; For instance, in the Airbus 380, humans no longer directly control the movement of flaps. Instead, the constant calculation of some computers dictates all their movements.

The pilot has had to get proficient at relying massively on the computer. Such changes made retaining the same method of training pilots impossible.

Pilots undergo two phases of training: with and without a computer

Originally, pilots received only training to “fly the plane,” as computers did not exist. Paradoxically, when computers were introduced, the peak workload increased as there were now three logics interacting, those of the human, of the computer and of the mechanical systems. The impact on accidents was dramatic, with human-computer interaction (HCI) becoming the crucial weak link of flying. Subsequently, the share of training dedicated to managing the computer had to increase drastically.

Now, this raises an obvious question: does it mean that those pilots can’t “fly the plane”? Well, it’s complicated… Modern pilots are taught to “fly a plane”. They have to understand the basic behavior of aircrafts and how they respond to various actions. Therefore, instructors teach initial training on a plane without automation, typically a small turboprop. Even experienced pilots benefit from continuing to fly in such a raw manner, for instance, through the leisure practice of gliders, which requires a very physical and direct relationship between the pilot and the aircraft.

Yet, a second phase of professional training amounts to learning how to fly planes through the computers. This occurs first in simulator, and then in the real thing, the very large or very fast jets that computer allows flying reliably.

Dual Mode Pedagogy in the Age of AI

The experience in the aviation industry points towards an upcoming evolution of pedagogy, AI becomes compulsory in the pedagogy of professions where AI matters in practice. For example, if practitioners use AI to conduct certain activities such as market studies, strategy development, or communication, then the pedagogy for those activities should eventually adapt to incorporate AI.

However, aviation’s evolution suggests a dual process by which the pedagogue delineates unassisted skills, on the one hand, from skills to master with AI’s assistance, on the other hand. Not only does the content drastically differ, but the evaluation too.

In the unassisted skill phase, one might have to enforce the elimination of the computer out of the process, for instance, returning to paper essays. In the assisted skill, one must ensure that a standardized computer environment is available for evaluation. Probably implementing this distinction will require separating the two processes to ensure full acquisition of both skill layers.

AI as an Obsolescence Threat to Various Management Skills

This dual mode of pedagogy aims to avoid a serious threat to current training: if a significant portion of the skills that used to be expected from graduates can be substituted by AI tools, neglecting the training of AI-assisted skills will likely disadvantage those graduates in the market.

For instance, consider the ability to produce dense planning of business processes, expressed through tools such as financial calculations in Excel and managerial explanations in PowerPoint. Such skills once proud markers of business school education. Students came to acquire them in the hope of gaining employment in consultancies or in management.

However, generative AIs excel at producing such apparently thoughtful content. We are not assuming here that AI is already above humans. We only need to observe that AI generates content that matches what evaluators used to expect from students: it is this very problem that triggered the outcry of many pedagogues begging that we remove AI from evaluation process!

However, solving this issue requires more than just rules and procedures. It probably indicates that machines could de facto substitute for what we used to teach. Hence, our pedagogical mix should respond to this new layer of technology.

The business school environment has often extolled the beauty of Schumpeterian creative destruction in the various industries that we study, like entomologists observing ants eating dead bugs. Whether we like it or not, it may now be up to us to stay ahead of the curve when a large share of our white-collar skill training meets the fast-feeding frenzy of the machines.