Generative AI and the creative confusion of academic writers

In his guide to productive academic writing the psychologist and writing coach Robert Boice distinguishes between non-starters and non-finishers. Even if there could be some academics who experience both problems in their writing, Boice suggests in his experience these are distinct problems with a different psychological basis underlying them. While “an inability to finish resembles its counterpart, an inability to start, in dynamics such as perfectionism” he observes that “the two forms of writing problems typically occur in different people”. He offers examples of writers he has worked with who are immobilised by the expectations they have imposed upon themselves, literally unable to get started without immediately comparing what they are able to write with what they feel they ought to write. They might get some words down but immediately stop to revise them, finding themselves caught in a loop of self-doubt which makes it impossible to proceed. He suggests that such writer are often “impatient to finish” if and when they manage to get started, relieved at overcoming their initial impediment and eager to bring the ordeal to a close. In contrast he writes that non-finishers “don’t struggle much over beginnings” because “they know that perfection comes later in multiple versions”. In fact some “non-finishers even write prolifically, without ever completely finishing their projects”. Their problem arises when a completed text has to be cast out into the world to be read by others, rather than the discomfort they feel when reading back a work in progress.

The point he is making is that perfectionism, ultimately a fear of failure, can be found lurking behind both of these challenges. The difference is a matter of where the perfectionism is located. For non-starters, the sense of what the writing should be makes it near impossible to get started. For non-finishers, the sense of what the writing should be makes it possible to bring their piece to a close. In both cases the writing is approached through a prism of expectation, imbued with expectations and obligations which squeeze out the possibility of enjoyment. If you constantly compare what you are writing to what you feel you should be writing then it is difficult to find enjoyment in the process. Rather than relaxing into putting thoughts into words, letting ideas pass through you onto the page, it will be a twitchy exercise in failing to live up to your own standards. The problem here is not having standards, as much how we relate to them. If they are a test you impose continually on yourself then you are, ironically, much more unlikely to meet your own standards. In denying the space to make mistakes, to fall short and to fail the possibility of growth and development is also lost. This is why once you get stuck in your writing, whether as a non-finisher or a non-starter, it can be hard to get going again. The more you see your work through a prism of your own imagined inadequacy as a writer, the harder it will be to actually write, let alone find enjoyment in the process.

However what it means to be ‘stuck’ transforms once you become familiar with the use of large languages models (LLMs) like OpenAI’s ChatGPT and Anthropic’s Claude. There are still difficulties inherent to the writing process which you routinely encounter: the paragraph you don’t know how to finish, the article you don’t know how to fit together or the deadline you are struggling to meet. The change arises because machine writing always offers a path forward. If you simply provide it with the rest of the paragraph, it will complete it for you. Claude suggested that I complete the paragraph you are reading by observing how this “newfound ease creates a profound shift in how we relate to writing – where being stuck once demanded we dig deeper into our thoughts or define our argument, we now face a constant choice between pushing through the difficulty or letting the machine smooth our path”. Not only did it complete the thought, it matched my style because I’d presented it with enough of my writing that it could plausibly write in my voice. In this case I used this to illustrate a point by retaining the distinct contribution of the conversational agent, framing it as an object which I could comment upon in order to develop my argument.

If I’m tired or stressed, working towards towards a deadline which I’m struggling to meet, this can be incredibly enticing. Over the two years I’ve been working with LLMs on a daily basis, I’ve noticed that I often keep working at a point where I would previously have downed tools and gone to rest. There are points where stopping would have been unavoidable, simply because you run out of energy and can no longer do what you were doing. Yet at this stage the LLM can now step in in order to keep you going by offering suggestions about a potential path forward. This can be incredibly dangerous in a sector already prone to celebrate, even demand, over work.

However if used in a more careful and restricted way it can be extraordinarily helpful. You know that feeling when an insight is on the tip of your tongue but you can’t quite put it into words? By sharing your sketchy ideas with the LLM it can nudge you into articulacy, helping you say what you couldn’t quite say yes. This capacity to crystallise half-formed thoughts goes hand-in-hand with the risk of working through fatigue. They are two sides of the same coin, in the sense that creative use of LLMs offers a way through what would have formerly been cognitive limits. This boundary-crossing ability raises important questions about intellectual labor and authorship. Where do we draw the line between helpful assistance and delegating our core intellectual work to machines?

There is a real and immediate danger that something important is lost if you are never stuck. Or if you see stuckness as a straight forwardly negative experience to be moved through as quickly and easily as possible. Creative confusion serves a vital function in scholarly work, forcing us to dwell with uncertainty until new connections emerge organically. This productive discomfort is what often leads to our most original insights, those moments when we transcend conventional thinking precisely because we couldn’t immediately find a path forward.

As Boice reminds us, “productive writing is much more than getting unstuck”. It’s not enough to simply move through writing problems, either on a single occasion or recurrently. To be a productive writer means building routines and dispositions which are sustainable and effective, rather than merely finding the way through the barriers which impede a particular piece of writing or writing in general. But if you remain caught by those barriers, unable to move forward, the possibility of being a productive writer remains foreclosed. The lure of machine writing lies in the diversity of ways in which it can help you become unstuck. Even if the full repertoire of means through which machine writing can help might not be available to an author, the simple fact of it being there as an ever present assistant designed to help has the potential to be an enormous moral, as well as practical, support.

My use frames it as an example for readers who might not be aware of how effectively frontier models can complete a paragraph. For those who are using these models on a regular basis, it’s no longer an intellectual curiosity to be observed but rather an immediate possibility inherent in the writing process. The relatively small numbers of sustained users of these models within the academic community creates an epistemic gap between those who are writing using the familiar apparatus of the 21st century academic writer (e.g. a word processor) and those who have incorporated conversational agents into their process. My experience has been that a significant shift occurs after months of routinely writing in dialogue with them, leading to an intense awareness of their capacities at points of intense difficulty.

The solution to the problem is not to engage in a humanist rearguard action to drive these emerging technologies out of academic life, but rather to affirm the joy which can and should be found in writing, as well as finding actionable ways to cultivate that connection in our writing practice. The problem is not GenAI itself but rather a soulless instrumentalism which relates to it through the frame of efficiency. Instead I suggest we can have a joyful relation to conversational agents, which sees this technology in terms of intellectual interlocution rather than machine writing.

#academicWork #academics #creativity #higherEducation #scholarlyWork #writing