The Question
University educators – teachers and program managers – face an uncertainty regarding the ways they recommend (and sometimes enforce) that students use and refrain from using Generative AI in their studies. Among others, the following question keeps coming up in my own teaching: Is the old slow and hard way to acquire established core virtues and capacities of a discipline still necessary, or can we flatten the path of students and allow them to offload the cognitively heavy lifting to GenAI? For instance, when teaching philosophy, we know a slow and hard way that forms core philosophical virtues and competencies in students: Let them struggle with complex, abstract, distant texts, try to write and talk about them, fail, fail again, fail better, again and again, give them feedback, coach and encourage them along the way, and gradually they form desired virtues and capacities, such as intellectual stamina and perseverance, tolerance for ambiguity and deep disagreement, ability to parse and survey complex conceptual and argumentative fields, etc. But maybe offloading parts of it, maybe the repetitive, boring parts, provides a convenient shortcut to essentially the same learning outcomes?
Note that this is an uncertainty about the way to acquiring the established virtues and abilities, not these virtues and abilities themselves. Regarding the latter, the current indicators point in a rather clear direction: Whether it is in software engineering (Kam et al. 2025) or in legal practice (Choi, Monahan, and Schwarcz 2024), the actual competence profile of senior persons in the field has not changed significantly so far; what might have changed is that companies have cut junior roles that had allowed graduates to finish acquiring these virtues and competencies because they see AI as a cheaper and comparably capable resource to fill these roles (Feng, Yun, and Yi Wang 2026). Therefore, for the time being, I suggest we can safely ignore this worry. The uncertainty regarding the way, however, is real. Should we encourage students to offload some or all of the hard cognitive work to AI, or should we discourage it?
One Basic Distinction
I suggest that a helpful way to think of GenAI use during studies is to draw a basic distinction between friction-increasing and friction-decreasing uses, and that friction-increasing uses are generally unproblematic, while friction-decreasing uses merit further reflection.
What is the distinction supposed to capture? In essence, it is about a basic fact of studying: The goal of studying is not production, but formation. This means that, generally speaking, offloading cognitive effort to AI during your studies is analogous to bringing a forklift to the gym. In the gym as during your studies, formation is the goal, not production (which is exactly the goal in the workplace, where efficiency and production are central, and where the forklift is well-placed). A recent study nicely illustrates this: Researchers found that offering unconstrained GPT-4 usage hindered students’ skill acquisition in mathematics, while a regimented AI tutor that refused to give students the answers, forcing them effectively to do the thinking by themselves, could mitigate the negative effects (Bastani et al. 2025).
This suggests that allowing or even encouraging students to offload the hard cognitive labor to GenAI is problematic. What is not problematic, in contrast, is to use GenAI to increase friction, for instance, by generating mock exam questions that target exactly the weaknesses of a given student.
A Lingering Worry
Perhaps you are now thinking that this confidence in the applicability of the analogy of the forklift and the gym to the offloading of cognitive work in studying is questionable, indeed problematic: Could it not be that the students still acquire the relevant virtues and capacities, just more easily? After all, the analogy to the gym and the forklift has its limits: While the forklift would directly and fully replace human muscle, the use of GenAI cannot completely replace human cognitive effort – or at least it does not have to: Humans still have to decide on what AI to use, to assess and verify the outputs, etc. And indeed, while the majority of the available empirical evidence currently suggests that friction-decreasing GenAI use is detrimental to learning, there are a few studies that suggest the converse (Abd Algani and Shini 2025). Currently, we have good reasons, but we are not certain that students really need to avoid friction-reducing GenAI use.
The Lingering Worry And Pascal’s Wager
The basics of Pascal’s Wager
I suggest that the situation and problem just described can be addressed using a method of thinking tailored to situations of uncertainty, which was pioneered in the seventeenth century by a philosopher called Blaise Pascal. Following is the gist of what is now commonly called Pascal’s wager (pp. 224-231 Pascal 1987):
Either God exists and there is a final judgment or there is none. We cannot know, but we must choose, or better wager, and if we are reasonable, we place our bet according to the expected utility, that is, the utility times the probability of the outcome.
- If we wager that there is a God and there turns out to be none, then we lose nothing. We live a virtuous life, have good friends because we are good friends and can die proudly (give this utility f1).
- If we wager that there is no God and there is no God, we have maybe an ok life (f2), and if there turns out to be one, then we face final judgment, which might not end well for us (f3).
- If we wager that there is a God, however, and there turns out to be one, then we will be very, very happy for eternity (utility: infinitely high).
| God does not exist (p) | God exists (1-p) | |
|---|---|---|
| Wager for God | f1 | infinity |
| Wager against God | f2 | f3 |
The core idea of the set-up is to see that one bet – wager for God – dominates the other: in the scenario where God does not exist but we still wagered for God, nothing bad happens, perhaps f2 – the utility of a godless life in a godless world – is slightly higher than f1, but it’s not even clear that it is (Pascal suggests that being a virtuous person and a good friend is not the most horrible life imaginable). However, if we wager for God and it turns out that God exists, even if the likelihood of that scenario is very small, it gets outweighed by the fact that the utility of that scenario is infinite. So, if we are rational, we should wager for God. While many of the details of this wager are still subject of lively debate (see the excellent collection by Bartha and Pasternack 2018), Pascal introduced an entirely new way of thinking about such uncertain situations, one that would ultimately mature into modern decision theory.
Applying the wager to AI and teaching
Now, how can this apply to AI in education?
Either we explicitly block students’ friction-decreasing use of GenAI, or we think that this is not necessary anymore and let students use GenAI also for friction-decreasing purposes. As teachers, we cannot abstain, we must choose the option that we think maximizes the expected utility for students. I assign both scenarios – the one where friction-decreasing use leads to failure to form relevant virtues and capacities, and the one where it doesn’t – a probability of 0.5 each. Given the current state of the cognitive offloading literature, that seems a fair estimate (note, however, that setting the probabilities here matters, as we are not expecting an infinite utility, as in Pascal’s original case).
- If we choose to wager against friction-decreasing GenAI use (in short: FDU) and it turns out that this is necessary for students learning outcomes, we win, so that would be great (I intentionally refrain from assigning numbers to expected utility, as this would entail a pseudo-scientific impression of accuracy; the specific labels assigned to the scenarios are of course inspired by the fantastic HBO miniseries Chernobyl).
- If we wager to allow them FDU and it turns out that this really does not hurt learning outcomes, we win, so that would be great as well.
- If we wager against FDU, but this turns out not to be necessary, then this would be somewhere in the middle: not great, not terrible: We would put the students through hardship that would not be necessary (not great), but they would still end up as qualified, able graduates, ready to make a mark and contribute to society (not terrible).
- However, and this is the decisive point, if it turns out that FDU inhibits their formation of the relevant virtues and capacities, we effectively form inept graduates, which is the one scenario a teacher should try to avoid by all means. So, this scenario would just be terrible.
| FDU blocks learning (0.5) | FDU doesn’t block learning (0.5) | |
|---|---|---|
| Bet against FDU | not great, not terrible | great |
| Bet on FDU | great | terrible |
Now, in analogy to Pascal’s wager, there seems to be one bet that dominates, but negatively: When we wager for FD GenAI use, but it turns out that this blocks students’ acquisition of the necessary virtues and capacities. I think this is the only scenario that is clearly unacceptable, and that all teachers are required to work towards avoiding. In contrast, losing the opposite bet is clearly less bad: This is the scenario where we bet against FD use of GenAI, but it turns out that it would not have hurt learning outcomes if we would have allowed FD GenAI use. There, students still end up as able, qualified graduates, ready to make a mark, but it turns out that there would have been an easier way. So, in sum, it seems that we should aim to block FD GenAI use until such time as there is really very solid evidence that they do not actually need this to acquire the core virtues and competencies of our field.
Take-Aways
Not all GenAI-use by students is equal. If they use it to increase friction, it seems clearly unproblematic. In contrast, AI use by students to decrease friction seems problematic from a teacher’s perspective: we would like to contribute to students’ formation of the required virtues and abilities, as they have evolved in our disciplines and allowed previous generations of graduates to flourish and make a mark in their careers. Even in light of the remaining uncertainties about the question whether they could or could not acquire these virtues with offloading to AI, a Pascal-style decision table seems to recommend opting for blocking the friction-reducing use of GenAI.
PS: This is pretty much an ongoing concern and thus work in progress for me – so if you have any thoughts on this, I would love to hear about it!
Leave a Reply