Are you using large language models like ChatGPT or Claude in your writing? The evidence suggests many academics are doing this, ranging from simple editing assistance to more extensive collaboration. But as these tools become more integrated into our writing practices, we need to consider not just what they produce, but what their use might be displacing.
It would be like, as the novelist Anne Lamott (1994: loc 240) puts it, “discovering that while you thought you needed the tea ceremony for the caffeine, what you really needed was the tea ceremony”. The end is what makes the means possible, rather than the means being something we pass through to get to the end. As she puts it, the “act of writing turns out to be its own reward”. Or at least it can be. I’ll argue that the pleasure which can be found in the process, the sense in which we can enjoy academic writing, emerges from the rhythms of the writing itself. It’s the flow of ideas, the interplay between the inchoate and explicit, putting things into words. When you need the caffeine it can be difficult to really inhabit the tea ceremony. Difficult but not impossible. It involves reminding yourself of why you’re doing this, returning gently to the moment itself rather than the pay off which you feel it is leading you towards.
My concern is that LLMs represent something akin to caffeine gum which immediately give you a bigger hit than the ceremonial tea was ever going to be able to. Why go through the ceremony to get what you need when you can have it now? Particularly if you can have it with more force and immediacy then slow immersion in the ceremony could ever offer you? If we see LLMs as a means through which we can immediately realise our initial ideas, we miss out on the intellectual struggle which helps us move from initial ideas to more fully worked out ones. As Lammot (1994: loc 24) puts it, “Writing has so much to give, so much to teach, so many surprises”. LLMs can deepen our reflexive engagement with this process, helping us more fully inhabit the surprises it offers to us. The problem is that they probably won’t if we approach them through a sense of the effectiveness they offer for meeting our goals: dispensing with the tea ceremony and taking caffeine pills instead, without ever fully understanding where the value actually emerged. If it was the process itself which was always the source of meaning and enjoyment, rather than the outcome, we have to be extremely careful with any technology which offers to accelerate or short circuit the process. “If you’re all about the destination then take a fucking flight. We’re going nowhere slowly but we’re seeing all the sights” as one of my favourite songs of the 2000s once put it (Turner 2006).
The rhetoric of generative artificial intelligence leads us to focus on the outputs which models produce. After all this is what defines the category: the capacity of this software to produce text, images, video and other forms in response to natural language prompts. This is particularly true within higher education where the capacity of students to generate essays has dominated discussion around LLMs and what they mean for the sector. In the process the prompts are reduced to a technical exercise, often imagined as a simple command or a strange form of coding which is undertaken in natural language. The problem with this view, which either neglects prompting or imagines it as a technical skill which can be taught, is that it obscures the literary quality of how we engage with LLMs.
To use models regularly involves a great deal of writing. To have satisfying and effective conversations with them involves writing thousands of words. Rather then a form of instruction, in which the powers of natural language and subordinated to some arcane computational logic, I suggest that we should approach models in terms of creative writing. I suggest it is literary in the sense of having a deliberate form which is intended to create a particular effect. It is certainly not an emotional effect which is being sought, but there’s an attempt to elicit a response from the LLM through a style of writing which has been specifically cultivated for this purpose. It would be a stretch to say that effective prompting is always good writing, but it will always be expansive writing. It will need to convey enough substance, illuminating a domain in which you are working, for the model’s response to be relevant to your needs.
If we see models as an escape from the messiness of writing, we risk losing something critical to our intellectual engagement. Even if the results tend to follow a rigid format, writing itself is rarely a linear process. It is often messy, frustrating and inchoate. It forces us to grapple with what we’re thinking through repeatedly trying to say if, often leading to many failed attempts before we get to something which feels satisfactory. The lure of machine writing lies in its capacity to immediately get us to where we want to go. The problem is how this disregards the journey, not just as an enjoyable process but as a formative one. Not only might we not know where we want to go to, the process of figuring it out contributes to becoming who we are. It’s often the mistakes and missteps, the wrong turnings which force us to grapple with difficult realities, which have a greater impact than those times when everything goes according to plan. Our meandering explorations often lead us to a place which a deliberate plan would have failed to. In writing, as in life. The point is not that we shouldn’t plan but rather that we shouldn’t let our expected destination exhaust the experience of the journey. We often, as Elbow puts it, need “to say the wrong words to get to the right words”.
If we immediately reach for the output, we either miss this crucial step or rush through it so quickly we fail to learn from the experience. What matters is the tea ceremony rather than the caffeine we consume through it.
Mark Carrigan 