Month: August 2025

The incorrigibility of ChatGPT and the end of teaching

Suppose you witness an accident and have to report it to the police. “Well, the red car drove with high speed towards me and then took a sharp left turn”, you exclaim. You try hard to find the right words to capture the precise sequence of events and people you’ve noticed. Being a bit fussy by nature, you keep correcting yourself. “To begin with, it seemed as if the driver wanted to run me over.”, is the formulation you eventually settle on. – Now imagine you try to refine your impressions using ChatGPT. Obviously, there is always room for improving on style and grammar. But can you expect ChatGPT (or any LLMs or AI) to improve on the accuracy of your factual statements? No. But what does it mean if it is used to that end anyway?

Given the way ChatGPT works, it has no experience of the world in any sense. Rather, it generates sentences by predicting the most likely sequence of words based on the input it receives and the vast amount of text it was trained on. Thus, it cannot improve on the content of any statements relying on experience. While this is no surprise, the repercussions it has for teaching contexts deserve careful attention because such contexts thrive on the correction of statements. Especially now that people not only use this device to hide their plagiarisms, but also to “decide” all sorts of questions. I’ve been wondering repeatedly what precisely it is that goes wrong in teaching contexts with the use of AI and now I begin to think that it comes down to a loss of corrigibility, a loss of understanding what corrigibility even means. Put in a nutshell, seeing that this device improves the form of (written) texts in amazing dimensions, it makes us blind to the fact that it impoverishes our relation to empirical content. In what follows, I’d like to explore this loss with regard to teaching philosophy.  

What is corrigibility? Corrigibilty means that a statement or text can be corrected. If I state that it’s raining you can correct me by pointing out that it’s in fact not raining. We offer and receive corrections all the time. We improve the way of phrasing something we’ve seen by finding more adequate phrases. We can differentiate between grammatical and stylistical corrections as corrections of form as opposed to content, but often the two are difficult to keep apart. The phrase “it’s raining” is formally correct when used among English language speakers, but what makes it correct for these users is how it’s applied to a shared experience of the world (in which it happens to rain). If I ask you to refine your phrasing, suggesting for instance that it’s really pouring and not just raining, I can mean at once to pay attention to your experience and the conventional way of expressing such an experience in the English language. When you think about your experience and modes of expression, you’ll likely involve linguistic sources (your language conventions, literature, the sociolect your audience is supposedly expecting etc.) as well as non-linguistic sources (whatever you can gather from other sense-modalities). Most importantly, you’ll involve relations of applying linguistic resources to non-linguistic experiences. In other words, we relate linguistic conventions to (non-linguistic) facts. ChatGPT, by contrast, doesn’t do that. Having no relation to the world, it is confined to linguistic resources; it has no other sense modalities and it has no way of relating linguistic to non-linguistic facts. In other words, while it can improve on formulations, it cannot be corrected. Put in Wittgensteinian terms, whatever seems correct to ChatGPT is correct – and that means that there is no sense of distinguishing between correct and incorrect. (There is an intriguing piece about learning this the hard way.) Thus, we shouldn’t even say that it’s “hallucinating” when it’s “making things up”. There is no meaningful distinction for this device between hallucinating and getting things right in the first place.

Now I doubt that I’m spreading any news here. So why is this worth saying? Because both the language of ChatGPT and of the merchants of this technology constantly suggests that this device is learning, being corrected and improved. Yes, it’s being improved at what it does already, but it’s not improved in any other sense. This lingo tricks many of us into thinking that the improvement is of the kind that we are familiar with. Just like AI is now increasingly taken to be a meaningful, sexy or caring interlocutor, it tricks many of us into assuming that it could “learn” by being “corrected”. But learning, for humans, always involves a relation to the world. The great confusion about ChatGPT, then, is that it would be improved in any way that we would try to improve our own way of expressing ourselves.

How does this affect teaching (philosophy)? There are many pieces about the decline of the humanities in the face of ChatGPT and related devices. Given how this technology diffuses our sense of authorship and our reading culture, I’m inclined to think that our whole way of cherishing text production and reading will go out of fashion and become a nerdy niche. Just like long electric guitar solos or keyboard solos, which seemingly were ubiquitous in the 70s and 80s, are now a thing for a few nerds on youtube. So as I see it, the problem is not that students are faking texts; the problem is that most texts are considered irrelevant. Along with the skills and features that go into their production. Being able to write good texts is already irrelevant in world where so-called leaders get by without even glancing at their briefings. But let’s stick to the current story. My hunch is that the loss of corrigibility ingrained in ChatGPT is the outcome of a larger trend that was clearly recognised in Harry Frankfurt’s On Bullshit as early as 1986: Once you realise that you can convince without sticking to techniques of truth-evaluation, you can disregard truth altogether. After all, the question is not in what way ChatGPT is incorrigible. We can figure that out quickly. The question is why are we letting ourselves be corrected by a device that is incorrigible.

But that’s a question for nerds. Mastering long written texts, let alone writing them, then, doesn’t seem to hold much of a promise for anything now. This is not just because students have incentives to fake their work; it’s because there are hardly any incentives to produce such work in the first place. Why do you need to learn to play the piano if you have keyboards with automatic accompaniment? Of course, you might get sick of their sounds quickly. But who cares if that’s all that’s on offer?

So again: the problem is not cheating; it’s irrelevance. Writing this, I feel like a fossil decrying the loss of its natural habitat. And that’s probably what it is: An old man whining that no one recognises the beauties hidden in the art he cherishes. So what? So what indeed?

So what’s left for teachers? If you don’t worry too much about plagiarized texts, you might adjust your energy towards getting people to think, not by by analysing texts, but by coming up with good prompts for ChatGPT or by enhancing your techniques of video editing. In other words, while certain products (such as well-written essays) will simply be done by ChatGPT in the future, you can support students in improving “their” work by focussing on helping them to use this and the AI devices to come as a good tool. The remaining question is, though, what this tool is good for, once we admit that writing texts is irrelevant?

What does it say? The supposed objectivity of written texts

“… interpretation is the source of texts, facts, authors, and intentions.”

Stanley Fish, Is There a Text in This Class?

Do you remember when you first committed some of your own thoughts to paper? Perhaps you kept a diary, perhaps you wrote poems or lyrics or crafted a letter to a friend. Perhaps you had worked on the aesthetics of your handwriting. Anyway, there it was. Something that you had written could now be read and, of course, misread in a distant place during your absence. This striking distance became even more evident to me when I had seen my words, not in my clumsy handwriting, but in the typeface of a word-processor. Imagining that someone would read my words not as my personal scribblings but as a text in an authoritative typeface, made me at once proud but also seemed to diminish my personal impact on the text. In any case, the absence or possible absence of the author from something written, I suppose, is what turns texts into something objective. As I see it, texts become objective when they can be read independently of the writer, of what the writer says and thinks. If this is correct, it seems that written texts are fundamentally different from spoken texts or thoughts. In turn, this makes me wonder whether it’s written texts alone that afford the interpretive openness allowing for different readings or interpretations as we know them in the humanities of our time. In what follows, I would like pursue some perhaps naïve musings on this issue.

Thinking versus speaking versus thought?

If you observe what you say in contrast to how you write, you’ll probably notice a stark difference between spoken versus written language. While academics sometimes seem to try and imitate the grammatical standards of their written language in their speech, we quickly notice that the grammatical rules, word choices and other aspects are vastly different. Pondering on this issue quickly brought me back to the ancient and medieval doctrine of “three kinds of language”, according to which thought is expressed through spoken language and spoken language is signified by written language. But once you notice how different already speaking and writing really are, it’s difficult to give much credit to said doctrine. The very idea that writing is a set of signs of what is spoken strikes me as a very impoverished understanding of the difference. This makes me wonder when written language was first considered as a set of signs independently from spoken language. Following Stephan Meier-Oeser’s work, my hunch is that William of Ockham and Pierre D’Ailly in their logical treatises are among the first to deem written signs as independent from spoken language. (Sadly, it’s not entirely clear why they hold this in contrast to many of their fellow thinkers.) Now, once you think of written language as independent from speech it seems that you acknowledge something that could be the objectivity of the written text. Of course, long before the written text is acknowledged as an independent signifier, there have been sacred texts like the Bible that were considered objective in some sense. But experiencing our very own writings as independent from our speaking must do something to the way we think about texts and their interpretability more generally, or so I think.

The written text as an objective ‘thing’

The way we encounter written texts or books (be it on paper or screens) seems to present them as distal objects, independent from how we interact about or with them. Like the table in front of you, the book on your desk or in your pdf isn’t altered when you look away. This experience is certainly at least in part responsible for the common assumption that texts and their meanings are stable items independently of us. Likewise, our experience of reading is commonly thought of as grasping something external to us or our interactions. But why? While I myself have begun to think that reading is in many ways a matter primarily dependent on interactions between readers, I equally wonder how written texts, non-sacred texts in particular, have earned the status of independent carriers of meaning that can be hit or missed. Our current reading practices inside and outside of academia seem to corroborate this assumption. – (What does it say? This is a question that silences classes but equally fosters the pretence that texts are stable unchanging sources of meaning that provide all the necessary constraints for possible interpretations. Yet, not knowing whether we’re reading a recipe or a a poem, we are probably unable to tell the genres apart without context. “Context” – this harmless little term obscuring all the greatly important factors allowing for recognition, and constantly underestimated as a “side issue” when it comes to competing readings!) But what does it take for a written text to be actually seen as independent in such ways?

Holland House Library after an air raid in 1940

The advent of ChatGPT

Investigating the question of the objectivity of texts will take some time. But currently it seems that this objectivity becomes undone in quite unexpected manners: the advent of chatGPT does not only call into question the production of texts through proper authorship. Rather, it also calls into question the independence of written language as a system of signs, thriving on a supposed text-world relation having been taken for granted for a very long time. Reading a piece of text, we can no longer presume that it was produced by a person having a relation to the world, to themselves and to other people making it a rational item, interpretable by rational beings, or simply readers.

How did we get here?