I attended an interesting meeting recently, where we debated the role of University and how our education should change in a world that is more and more taken over by generative AI. Since I need to write down a few words on this to share with various colleagues, I may as well do it in a blog post. This post integrates my thoughts with ideas that I picked up from the conversation. Where possible, I tried to look for and provide the appropriate references.

Yes, this is again about generative AI. If you are overwhelmed by the noise about this topic, I hear you, and I would recommend closing this tab and do something more enjoyable for today. Also note that, unless otherwise specified, I am going to do a disservice to the field and use AI to specifically talk about generative AI, and in almost all cases with this I mean text-based large language models more than any other media.

To begin with, I was caught by surprise by a strong push from a sizeable portion of our students towards an embrace of generative AI to replace skills that are deemed no longer relevant, like writing essays, summarizing content and, partly, even having a human tutor. It surely helped to set the stage and the tone for the discussion, and to highlight the increased disconnect between students and teachers, and the need to keep the dialogue alive and to make an effort to understand each other.

The intense marketing around AI is slowly creating a push to move from teaching students to collect and reflect on information to teaching them to critically evaluate AI responses and AI-provided sources. But if students don’t know how to find trustworthy sources themselves and how to filter the signal from the noise they contain, how can they critically evaluate what AI gives them? Even more so when we know that AI already embeds bias in its responses [SciAdv25], including source selection, and that it is in fact not that good in summarization [BBC25, TUE26].

But the marketing is working, and some Universities have already started to go all-in [Wonkhe26]. There are reasons, some of which we will explore below, that highlight a need to change some of the things we are doing as academics and educators, or at least how we are doing them. But I think going all in cannot be the solution, just as pretending that nothing happens is not.

This is prompting a few questions

What skills do we need, and how do we train them? Can we do it in another way? What do we actually contribute as a university?

There is value in avoiding universities to become obsolete and before even attempting an answer, I think it is important to put all the cards on the table and trying to read through the hype and marketing. Especially now that Venture Capitalists start pushing for a Return of Investment and the real unsubsidized costs of the whole operation start surfacing [Zitron26].

# Looking at the wallet

And this is the first issue. Asking unlimited token usage for lecturers and students, even just from a practical stance, is going to be increasingly expensive. Now that AI providers are changing their enterprise licenses to token-based billing, prices are skyrocketing up [Zitron26]. This should not come as a surprise: it was well established that premium subscriptions hide an order of magnitude of costs under the rug. For instance see the estimate in Willison26: $2,180.16 worth of tokens for $200.

These costs are already having a direct effect on company expenditures, see Willison26-1, or the many many links in Zitron26]. To the point that a company recently found itself wasting 500.000.000$ (yes, millions, you are reading it right) because they did not think of capping token use Axios26.

So whatever we do, we need to be mindful of the possibly enormous costs that are hidden in an enterprise or educational use of the tools.

# Looking at the research

At the same time, research already shows major concerns on many fronts.

The first is psychological well-being. Using AI in everyday work, even when you know how to use it, results in a high cognitive overload from the constant high attention needed for the critical evaluation work [HB26, Forbes26 both paywalled]. So this is not actually something that decreases what is needed of staff members, it in fact asks for more [HBR26-1, Forbes26-1, Forbes24 all paywalled], eventually leading to a high risk of burnout from all of these questions.

This is also being reflected in the increased amount of slop in research output. Recent studies that just came out are showing that the submission rate of publications went up by 40%, and the quality of publications is going down [Phys26, OrgSci26]. Again this is not just a matter of whether we will be able to do more with each other and adapt to what is needed.

In the end is up to us to decide which values determine our choices, and I would argue that human flourishing should remain at the center of it. This means that we should not ignore the worrisome effects of some of these choices on our health as individuals and the health of our fields of research.

For the review system, the situation has already become more unmanageable and unhealthy and also open research services like arXiv are struggling under the load of these changes [404-25, Futurism26, Science26].

# A broader look at ethics

And of course this leads to the often overlooked ethics discussion. It is very easy to “waterwash” all ethical concerns by asking about the ethical implications and water consumption of watching a netflix video or making a google search. But this is just a decoy to avoid talking about the deeper problems behind our current AI gold rush:

• Data theft: most models are trained on stolen, copyrighted data [EmpireOfAI].
• Hardware waste: 70% of data centers are unused, yet companies race to build bigger ones.
• Regulatory abuse: Companies like Grok pollute regions with impunity because they’re too wealthy to be touched [Politico25, EmpireOfAI].
• Environmental issues: I don’t know were to start [Fortune26, NYTimes25, Atlantic26, EmpireOfAI, PulitzerCenter25]
• Exploitation: Underpaid workers in developing countries work with tagging or filtering data used to train AI models, often with bad mental health consequences. [Arte25 - reportage remains available until 2028, EmpireOfAI]

An in all of this, these massive datacenters, rushed to accumulate computational power and the promise to install AI helpers everywhere, remain on but idle 70-80% of the time [Fortune25, LinkedIn25 many references are provided in this last, less official, link], and I am not even trying to enter in the jungle of privacy and security issues connected to the use of AI tools and agents.

I think students with ethical concerns have every right to refuse AI, and just as vegetarians refuse meat and we would not force them to eat a steak in class, I feel very uneasy to impose the use of AI to those students.

And yes, I am a hypocrite, coming here to preach the ethical issues and still being an active user myself. It is a duplicity that I am still struggling with. But whatever we choose to do, we should not ignore these issues.

# Back to the job market

Let’s assume now, for the sake of argument, that all of the above can be ignored and we want to train the next generation for a new job landscape. What will that landscape look like 3-5 years from now, once they leave university?

I don’t think we can now. The market embrace of AI has been oscillating wildly, with Duolingo and others going fully automated to only backtrack on their steps shortly after [Inc26]. There are many reports of most pilots with AI at companies failing [MITReport25]. A lot of the layoffs are happening at tech companies and are usually an excuse to save or shift money to other investments [Guardian26, CNN26]. OpenAI and Anthropic themselves have started shifting the narrative away from AI replacing all jobs, to AI replacing some tasks [BusinessToday26]

Months ago the message was that companies would not hire juniors anymore, to only then realize that once the seniors left, they have no one to replace them and the cost of training a junior developer from scratch, without the expert guidance of a senior mentor, will start adding up both in time and money [Dubach26]. In the meantime, like it has happened many times in the last few months, they just “learn” with Claude Code and suddenly wipe-out their un-backup-ed production databases…

There are also the prospects of outages in the services, the AI bubble bursting, another conflicts or the US and China pulling the access to their services. What happens in that case? Who will be useful if we have only trained a population of prompt evaluators?

# My hopes

I think we should be mindful that the environment evolving fast also means a large degree of uncertainty. In particular, also the interaction between companies and AI, especially with the amount of money poured into the marketing, is changing fast. And we cannot just look at what is happening now to make a decision, or to predict what will happen in the future.

Does it mean that we need to sit still? I personally don’t think so. What I’m arguing is that we should follow at least the lead of research results, to try to align our decisions with what we know is working or not working, remaining mindful of the dangers.

And sometimes this will mean that we will keep waiting on some changes until it’s clear what makes sense, or at least we will be the one making the experiments. But not as a blind, full-in buy-in, I mean a proper experiment, with careful monitoring, contained reach, where we can see the outcomes and evaluate them. And if it works, good for us. And if it doesn’t, we can take a step back, present our results, and try a different experiment.

So that’s what I would be advocating for: to be flexible to maintain a view on the future, understanding that things are changing so fast and so that we don’t really have a clear target that we have to aim for, and experiment on the possibilities by maintaining a careful eye on the research, avoiding a full embrace until it is clear that it is beneficial to us as people and society.

# Education and the university

To what extent can we really have a common vision and a common way of working here? To what extent must there really be differences between disciplines in how AI is used?

I think a common vision is possible, but requires clarifying our values and our priorities up front. Recently a friend and colleague pointed out that this is very clearly expressed in the recent encyclical Magnifica Humanitas. I just copy here her quote from the letter:

We are called to reflect on the great “construction sites” of our era and ask: What are we building? As technological development rapidly transforms languages, relationships, institutions and forms of power, we (…) must and can choose which projects to work on and in what manner, so as to safeguard and value the grandeur of humanity that has been given to us as a gift.
(…)
It is not possible to provide a single, comprehensive definition of AI. What can be stated, however, is that we must avoid the misconception of equating this type of “intelligence” with that of human beings. These systems merely imitate certain functions of human intelligence. In doing so, they often surpass human intelligence in speed and computational capacity, offering tangible benefits across many fields. Yet this power remains entirely tied to data processing. So-called artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships and do not know from within what love, work, friendship or responsibility mean. Nor do they have a moral conscience, since they do not judge good and evil, grasp the ultimate meaning of situations, or bear responsibility for consequences. They may imitate language, behavior and analytical skills, or even simulate empathy and understanding, but they do not understand what they produce, for they lack the affective, relational and spiritual perspective through which human beings grow in wisdom. Even when these tools are described as capable of “learning,” their way of doing so is different from that of a human person. It is not the experience of those who allow themselves to be shaped by life and grow over time through choices, mistakes, forgiveness and fidelity. Rather, it is a form of statistical adaptation based on data and feedback, which can be very effective, but does not imply inner growth.'

Learning is a deeply human and social act. It requires a careful selection of the material and the exercises from the lecturer, in a way that fits the class, its positioning in the curriculum, the available time, and the relevance in the students’ development. This has to be adapted dynamically for each cohort of students, listening to their concerns and struggles, and finding the common way in a two-way conversation that requires a mutual understanding. An LLM can only simulate this and hallucinate a response that looks appropriate, but more often than not misses the nuances and the larger picture.

This does not mean that it cannot find a useful place in various part of the process, but I think it is dangerous to think it will become an integral and central part of it. Especially when there is growing evidence of the harm it brings in the development of critical skills [Berkeley26, Mit26, Hartman26, Futurism25]. Not to mention the fact that prominent studies showing positive impact of AI in education keep being retracted [ArsTechnica26].

I agree that more and better can be done, as for instance advocated by Eric Bachman, to promote AI literacy and provide an interdisciplinary perspective on the ethics, psychology and science behind AI. I think we are beginning to do this more and more, and our University has widespread expertise on all the aspects directly available.

I’ve also been saying for a long time that I hope that this disruption will be the chance to revise the way we teach, shift the focus on promoting the learning process rather than the final grade. But this doesn’t mean that we have to change everything else to do it or that we should buy-in and deskill our students by “letting ChatGPT do it” and shift the focus on revising its answers. Even if they have the impression that this is what is asked from them by society. It is also part of our role to cut through the noise, push back against the nonsense and steer the discussion to a healthier place. And in this case, we want to ensure that whatever we do, AI will serve education, and not the other way around.

So let’s find the middle ground where we are not giving up the important things, we promote literacy and critical thinking, and we don’t give up in shaping elastic minds that can run the world of the future.