A Few Theses on LLMs ("AI") at the University

In the summer of 2021 I attended a workshop on BERT and GPT-3, two models that were keeping computational linguistics on its toes. A new model architecture from Google showed no signs of a performance plateau as the amount of training data increased. In other words, more was better. Millions and billions upon billions of pieces of text were fed in, and the model got linearly better the more it was given to read. "Better" here means that it could fulfill its one and only task, completing text, ever more coherently and faithfully to the meaning of the words. In the workshop we considered what the future impact of this technology might be. As linguists we saw above all the benefit of getting behind the surface of words directly to meaning; that way one could better investigate how grammatical structures interact with meaning, and do so across hundreds of thousands of texts. I wrote my master's thesis on singular 'they' and on how this GPT-based technology might one day make up for missing empirical knowledge – in the thesis itself I did not achieve the performance I had hoped for, but the potential was obviously there. Apart from that, opinions differed as to what would happen. My assumption was that the technology would only be used as powerful text completion; that is, for routine texts that mostly sound alike anyway and require no creativity. Just as weather reports were the first texts that software could translate perfectly, I was thinking of repetitive press releases, of advertising prose and of political language.
In my master's thesis, however, I warned against classifying the language of the GPT models as 'intelligent.' In the ethics section I wrote about how mapping linguistic proficiency onto intelligence is often unjustified, how it disadvantages people with language disorders, and how it could lead the majority of users to mistakenly attribute independent intelligence to a talking robot dressed up to serve a purpose. As a staunch anti-Whorfian, I found this way of thinking deeply repugnant. Yet half a year after my thesis, OpenAI's marketing had already confirmed my fears: artificial intelligence has been the battle cry ever since, and the companies deliberately fuel the misunderstanding.
Of all sectors, it is in my view the education system that is struggling with this the most. Teachers speak of a "lost generation" that the hype around the new technology is leaving in the lurch and causing to miss important steps in learning. What mostly comes out are mediocre to poor performances that nevertheless read better than what teenagers and young students are normally capable of. Nobody yet sees the way out of the crisis clearly, me included. I did warn about it very early on, but I had no solution then and have none now. In what follows I would simply like to put forward a few theses on the subject of large language models (LLMs). They come on the one hand from my own experience and on the other from a keynote lecture by Professor Mark Dingemanse. I consider Dingemanse the most competent thinker in this field; he manages to say truly original and useful things on the topic.
1The Difference Between Expression and Term
Precisely in this context it is important to become aware that an "expression" and a "term" are not the same thing. An expression is the linguistic rendering of thoughts on the surface; it concerns the manner of saying something. A term, by contrast, is much more narrowly defined. A "term" certainly has linguistic form as well, but behind it there is always a concept or a concrete idea. Behind the term, moreover, lies a longer discourse within a community, a scientific discourse, say, or a societal one. There are various expressions for the fact that Switzerland often keeps out of alliances and conflicts internationally; neutrality, in turn, is a term that has been negotiated over generations by different voices. Another example is academic prose: to outsiders it seems unclear and turgid. In academia, however, the rule is that the expression must be clear while at the same time you must use all the relevant terms in order to tie in with the preceding discourse. Academic terms often seem heavy and, because of their age, a little dusty as well; but they perform the important task of placing a text in its discursive context. That is why precise word choice is so important in non-fiction texts.
And this is where LLMs fail especially often. Their inner structure recognizes, purely statistically, that academic texts have a vocabulary particular to the genre, and they can imitate it quite well. The sheer effect of an 'academic' text from the pen of an LLM is convincing. But the LLM fails to distinguish term from expression and therefore uses terminology that is completely inappropriate to the field. Every second LLM text on linguistics, for example, contains the words "linguistic landscape"; it sounds nicely academic, but it actually denotes a specific research approach in sociolinguistics and is therefore rarely in fact appropriate. Only those who know the respective discourses notice such things quickly. Those who don't have no chance and – this is the serious part – cannot learn anything useful from such a text either.
And that is why it is enormously important to read and learn independently: you have to make yourself independent of the linguistic expression and understand the discourses of content. Only then can you also communicate effectively with your community.
2Prototypes Take Over
Roughly outlined, prototype semantics says that concepts are assigned to categories, each category having a central example. A sparrow is the most prototypical bird, the eagle is still relatively typical, the crane rather less so, and the poor penguin is only just allowed to be part of the category 'bird.' This way of thinking is inherent in humans in many contexts; thinking in prototypes thus leads to stereotypical assumptions such as the old white boss and the foreign cleaner – the harmful potential is obvious. It is, however, precisely a strength of human thinking that the mental categories are nevertheless inclusive and that peripheral elements and patterns can still be classified correctly. It is also, I think, a strength of the academic world that you are constantly urged to go to the periphery of the categories.
By their nature, LLMs could likewise be made to write sometimes more centrally and sometimes more peripherally. The commercial AI companies, however, have little interest in the latter. It is too uncertain if the LLM were now and then to produce surprising output. With millions of requests every day it is a risk if a thousand of them take a gamble. That up to ten of those will then also contain something unpleasant or harmful is statistically very likely, and the company is liable. Accordingly, LLMs are statistically conservative scribblers. In my experience this shows in a tendency toward excessively 'prototyped' expression. While humans do share this tendency, they can also deviate from it well; if an estimated 80% of people would answer a question with prototypical approaches, there are still 20% who deviate and keep the discourse alive. With LLM answers this split lies at almost ~100% prototypical and 0% peripheral output. That yields no intellectual progress and no intellectual diversity. The societal stereotypes mentioned above are then also hard to avoid.
In this context Mark Dingemanse quotes the artist and music professor Enongo Lumumba-Kasongo: "[creativity] is so much about trying out weird things — and AI is so much about averaging. It's the people invested in playing in that space of the anomaly and playing with the unexpected who will continue to thrive." – not only do I very much agree with Ms. Lumumba-Kasongo, I think many people feel this need now more than ever; and it is a glimmer of hope, and a possible sign of this need, that art such as that of Angine de Poitrine can take the indie music scene by storm.
3You Already Know the LLM's Answer
Here I am relaying theses of Prof. Mark Dingemanse's which, to my knowledge, he has not yet published. He compares the reception of LLM output with the reception of horoscopes, palm readings, tarot and so on, and with Garfinkel's experiments. The latter were experiments in which test subjects spoke with a supposed therapist who, however, only answered their questions with "yes" or "no." This "therapist" was reading from a randomly generated list of “yes” or “no”; the answers were thus completely arbitrary. And just as the lines on our hand or the cards laid out are in themselves random patterns, they evidently still offer people a useful surface onto which they can project meaning. For Garfinkel's subjects were very much able to get something out of the conversation with the yes/no voice, as long as they believed there was an intelligence behind it. And it is plain that esoteric fortune-telling is practiced primarily for emotional reasons. And it is also clear that the only possible source of the conclusions drawn is the person doing the reading, and not the cards or the horoscope. In short: the arbitrary patterns make you conscious of the preconscious.
LLMs, according to Dingemanse, have the same or at least a very similar effect: the pleasure of reading the output feeds on the same curiosity that also makes horoscopes seem so interesting to us. The illusion of the intelligent LLM is probably that we are seeking to confirm our own ideas there. This effect also grows stronger because the commercial AI companies train their new models on interactions that were rated positively. LLMs are thus continually turned into models that say what one would like to hear. And then it holds all the more: the insights gained are not new, for their source was, after all, the user – only the user is mostly unaware of it. Set against this, research papers seem like ungratefully unwieldy monstrosities that give you no quick reward. That is not pretty, but unfortunately it is necessary for valuable work.
4If You Don't Know the Answer, Don't Ask an LLM Either – Learn It Yourself!
The converse of the last thesis is simply that an LLM should not be consulted when you do not know the answer. Or, put differently: when you cannot evaluate the answer yourself. As briefly touched on above, usable texts require knowledge of the underlying terms as the results of discourse in a field. Knowing the discourse also includes understanding exactly what the conventions and requirements are of the text you are writing. The output of LLMs can therefore only be used as long as these conventions and requirements are met. But for that you have to be a capable reviewer. And if you are one, the task of LLMs is reduced to merely saving time; for you could just as well have written the text yourself.
The use of LLMs in an academic context is therefore problematic. The whole point of graded assignments is that a trained person shows you where the weaknesses of the text you produced lie. I suspect, however, that the learning effect only sets in when this feedback refers to your own text. If you have handed in LLM text as your own, the feedback does not concern you. Accordingly, the experience of submission and feedback is also valued differently on an emotional level if you have not gone through any effort and do not personally stand behind what was written. That, too, I suspect, is ultimately harmful to the learning effect that the university is supposed to produce.
5Independence from LLMs Becomes Valuable
"If you don't use AI regularly, you are going to have massive problems in the next few years. Both professionally and privately," an acquaintance of mine remarked. We see similar ideas in the frequently repeated narrative: "AI will not replace human jobs; people who use AI will replace those who don't." I thought so myself at first, since the idea certainly makes sense: a useful tool should be learned and not ignored on principle. My thesis at present, by contrast, is that this time is already over before it has really begun. From 2021 until roughly the beginning of 2025, LLMs were still much more sensitive to prompting language and strategies. On online platforms people exchanged tips on how certain tasks could be solved through targeted prompting. It is exactly this circumstance that the commercial AI companies have since tried to correct: the so-called reasoning models are ones trained to prompt other LLMs in a targeted way. The results did in fact become better and more independent of the prompts. The human being as 'prompt engineer' is disappearing from the equation – or has already disappeared.
What is certainly true, though: LLMs can take certain routine tasks off our hands and in that way make us more efficient. If we can assess the output well (see above regarding terms & reviewers), then we can also use LLM output for our purposes. To that extent the back-of-the-envelope reckoning makes sense that the worker who uses AI is more valuable than the one who doesn't.
The individual, however, becomes valuable not through mere efficiency but through what sets them apart. It often comes down to possessing skills that are, relatively speaking, rare and that thereby make you unique. And in a world in which everyone has access to a reasoning model that can prompt itself wonderfully, the individual's aim should accordingly lie in skills that lie beyond that. It is of course difficult to say what exactly those will be. But it will be important to be able to assess LLM output from the perspective of an expert. And for that, a certain independence from the LLM is indispensable. So I contradict the narrative by deliberately motivating people to develop their skills without LLMs. For then they truly have something in hand that distinguishes them from mere prompters.
6You Have to Understand Things as a Whole
It is difficult to formulate this insight as a catchy thesis. So here are two examples: in a text, a student referred to a scholar whom he quoted by way of a few terms. By his own account in the oral exam, however, he had never heard these terms. Worse still, the scholar had indeed made statements that seemed meaningful in context but had not used the wording quoted. The distinction between expression and term from above applies here, too: these were only supposed terms. But the truly problematic thing about the student's behavior was this: he had obviously not engaged with the scholar at all. She is a researcher from Ethiopia who is successful in the Western academic world and deals critically with the effect of 'AI' on how ethnicities are viewed. That is an interesting biography, and it is surely worthwhile to look beyond the rim of one's own topic to get to know her and all her theses. I am reminded here of my time as a high school student, when I researched the author of a particularly fitting article and even contacted him on Facebook because I found his work so interesting.
Another example concerns a student who had carried out a programming project. I noticed the use of regular expressions (RegEx) and asked him in the oral exam what he could tell me about them. He knew only exactly what he had used RegEx for, namely isolating a sought-after pattern in a URL. But RegEx are versatile and can be used for manifold tasks that have to do with patterns in text. Ideally, the learning experience of the project should have sent him on his way with this piece of knowledge. Then, in his next projects, he would have a useful tool in hand that he could apply flexibly.
The problem with learning with LLMs seems to be that you get to know the individual elements only in the context of the ready-to-serve solution. The incentives are currently set in such a way that only this very solution is of interest; you are practically finished before you have learned anything new. The value of an academic education, however, lies in getting to know the individual parts each as a whole before you put them to meaningful use for your own purposes. That is how you collect tools for later tasks and effectively broaden your own horizon. Both students would only have successfully learned something useful if they had the quoted scholar and the library they used 'ready for use' in other circumstances as well, possibly for other purposes. As things went, however, all they have in their heads afterward is what they already knew beforehand and had told the LLM as a prompt. Little that was new was learned, if anything.