In recent months there have been at least two controversies relating to AI-content and literature. In the first, Olga Tokarczuk, the Polish Nobel Laureate, said she used ChatGPT for her creative process (“How could we develop this beautifully”), noting “that in literary fiction this technology is an advantage of unbelievable proportion”, before having to walk back some her statements later following online outcry, stating AI was only ever used by her for research. In the second controversy, Granta published a prizewinning story which it seems may have been written by AI, and which may even have been reviewed by AI too. (Granta was neither prize-giver nor reviewer here)
These controversies made me realise that I had a different view from most commentators on these topics. At the time I considered AI, by which I mean here large language models (LLMs) largely deployed as chatbots, to be a neutral or possibly even positive tool for fiction writers, indeed for literature as a whole. At the same time, I acknowledged that it seemed obviously a negative thing for readers.
This divergent impact of the technology is a serious challenge for literary production and consumption, one which I am on both ends of as writer and reader. I wanted to work out if giving the topic some proper thought might allow for a more harmonious view of LLMs’ impact, and this blog post is the result of that initial thinking.
But before we begin, three things. One, these are not deep convictions. I would be happy to change them if I read or heard something that suggested I should. I have already changed, and nuanced them, as I wrote and revised this piece. Two, while this post is certainly more “pro-AI” than most voices on the topic, I do not use AI while writing my blog and have no intention to change that. Nor have I yet used it for my fiction. Here, now, you are reading me. My posts are proofread by my partner, who is like me a human being.
Finally, my focus here is on the use of AI by writers like Tokarczuk, rather than attempts to create an entire work of art through prompting alone. In other words, it concerns the writer’s possible augmentation rather than their substitution. I think this is a more interesting area, because the latter remains for now a relatively distant prospect. The former, meanwhile, demands our attention right now.
The Reader’s Perspective
For the reader, I would identify the chief problems with the involvement of LLMs in fictional production as centring on three interconnected topic areas: authenticity, dignity, and complicity in the systematic harms associated with data centre buildout.
Authenticity
While in and out of the ivory tower we might talk about the “Death of the Author”, readers in general want to feel they are reading a real person’s work. The disregarding of authorial intentions in interpreting a work is not the same thing as saying the author does not matter. Their individuality does not, but their abstract personality does. We want the words to feel chosen, otherwise the act of their interpretation stops seeming worth the time. Hence the relative lack of literary interest which the formulaic backs of milk cartons offer. The best books have the greatest sense of chosen-ness behind their words, whether that comes from the painstaking craftsmanship of a Flaubert or the abstract workings of genius that produced Dostoevsky’s more hurried, yet no less haloed, works.
Perhaps this is why ghost-writing appears a lesser crime than the use of LLMs. Even if it is obvious that this political or sporting figure hardly picked up the pen except to place their name upon the cover, we know that someone did. This is also why the overzealous editor, such as in the famous case of the editor Gordon Lish’s treatment of Raymond Carver’s stories, where Lish chopped and chopped and chopped at the texts until we readers ended up with the bare and brutal masterpieces we read today, is upsetting to those who know it without fundamentally compromising the texts themselves.
Perhaps Carver’s authenticity is compromised, but the texts’ authenticity is not – they still stem from a human source. In a similar way, I wrote in an earlier draft that newspapers and social media compromise authorial authenticity by placing heavy pressures upon them to write a certain way. But the end products in those cases are still texts produced by people. Compromised people, confused people, but nonetheless human beings like you or me. They are authentically human documents, reflecting human emotions and human struggles.
Dignity
Depending on the degree of AI infiltration into a work’s creation, of which more below, that’s not necessarily the case with a text today. A work whose words are not entirely an author’s own is no longer an entirely human document in the same way. This matters to many readers. Its importance is not at the level of the text, however. Students often read unseen passages and may well be unable, if not already now then at least soon enough in the future, to differentiate between human and hybrid or even machine-generated documents. Yet once we return to the document as a whole the distinction is more important. The problem with what Tokarczuk may have done, if I understand people’s complaints correctly, is that she denigrated herself as a writer. If you do not respect writing enough to want your own to be your own, the argument goes, you probably should not do it. And if you do not respect writing, why should a reader waste time reading you?
This argument does not have anything to do with the output upon the printed page, but about the context surrounding it. In that sense, it’s a moral accusation, a purity test. But it is one that is very different from being reluctant to read writers who were “bad” people. The alcoholics, the violent, the cruel – literary history is filled with such writers, and ever being replenished in its stock. We regret that Rimbaud stopped writing to be an adventurer, or that Nietzsche’s syphilis meant he was no longer able to write (in a way that made sense), or that Lermontov and Pushkin thought duelling for temporal honour was better than surviving to put more pen to more paper and buttressing their immortal honour.
Such failings fill us with regret for the missed works and unhappy families, but do not change what we have, those eternal masterpieces that fill us with gratitude for their creation. LLM use can make for an entirely different reaction. Usage does not reflect life getting in the way of art, but a kind of disinterest in art. If you don’t want to do the research for your book, why bother writing it? The implicit view behind this is that writing works of literature must be a holistic activity, where the artist must do all of it or none of it. It is a shame for a writer not to do something, not merely a moral failing. To find the traces of LLMs in a novel suggests a flourishing forestalled. The writer could have attained a more full writerly potential through efforts that have instead been replaced by a machine.
This is why the books can seem far more tarnished by association with LLM use than by an author’s personal misbehaviour. The latter means we get fewer books, but the former means we get less humanity in the books. It is quite fair, given this, for a reader to turn away, and refuse to read.
Complicity
The last set of reasons why LLM use is unpleasant for a reader rest upon the assertion that LLM use in itself is immoral because of the associated data centre buildout and therefore should be avoided as far as possible. These arguments are not so convincing to me, but they are important to others, and so I am setting them down here.
AI is driving a significant global infrastructure buildout through new data centres and the power systems needed to support them. Mothballed power plants, including those fuelled by fossil fuels, are being brought online provide electricity. New power plants are being built too, many using particularly energy inefficient but rapidly deployable technologies, such as the reciprocating engines and aeroderivative turbines used at xAI’s Colossus plants in Tennessee. While it is true that some of the new capacity being built is renewable and there is a resurgence of interest in nuclear power plants, gas is the primary fuel source due to the operating profiles of data centres. Gas is a fossil fuel. If you care about avoiding unnecessary carbon emissions, you can easily make the case that writers should write the old fashioned way. (The moderately old-fashioned way that is – by themselves, with a laptop or pen and paper.)
A similar situation obtains for water use. We live in increasingly water-stressed environments and should not contribute to this with avoidable LLM use. But just as there are different ways of producing electricity, there are different ways of cooling large industrial enterprises. While closed-loop cooling essentially limits the water use of a data centre by reusing it repeatedly and air cooling uses big fans to achieve the same result, other cooling types can see more wastage and impact on end consumers. (More details in this article from The Atlantic.) If you do not want to contribute to poor use of water, you can argue similarly that writers should avoid LLMs, and you should avoid those who do not. But this water argument, though popular, is weaker than the carbon argument. Near-to-mid-term economic gas alternatives are basically unavailable for data centres, whereas this is not the case for wasting water.
Other problems with the data centres themselves involve their impact on local communities. Power quality decreases (think flickering lights), and then there are things like headaches and other health issues. As with oil and gas production and refining, it seems that data centres are creating sacrificial zones around them, where society supposedly accepts local harm in exchange for global benefits. As with mining and oil and gas, however, there is the problem of environmental justice. This is the issue that such “local harm” tends to be concentrated in areas where people are less able to defend their interests, being poor or of minority groups. If you care about this local impact, you would be warranted to be critical of a writer who used LLMs thoughtlessly.
Finally, there are also impacts on local and regional power prices. But while it might seem as though data centres are entirely negative in their impact here, matters are likely more complicated, so it is perhaps best to focus on the other factors above, especially when their negatives are clearer. (For details on power prices, see section 8 in the International Energy Agency’s “Key Questions on Energy and AI”)
The above are general problems with data centres, ones that we have to think about personally every day unless we consciously add an “-ai” to all of our Google searches and are lucky enough to have jobs where there is no pressure to use the tools. More specific to writing, however, is the matter of copyright abuse. There are ongoing lawsuits against Meta and others over their use of literary (and other written) creations for training purposes without the consent of the creators or their compensation. A writer using an LLM is showing a certain lack of solidarity to their own people, and a reader who wants to show solidarity to the creatives displaced and stolen from by LLMs should not read those who are blasé about the tools’ impact.
The problem with these arguments, except the last one, is that the focus is on our and the writer’s contribution to the bad thing, rather than on the work. This allows the counterpoint of scale to be made. Your average programmer’s agentic workflow is several orders of magnitude worse than a writer asking an LLM a question – for the climate, for water, and for local communities. It becomes too much a matter of purity, rather than possible harm mitigation. I myself find only the dignity, authenticity, and writerly solidarity points convincing here.
The Writer’s Perspective
The previous section was an attempt to explore the arguments about why writers should not use LLMs from the perspective of a reader. The voices against LLMs are significant, strident, and seem to be growing in importance. Yet the problem with all of these arguments is that the closest they get to the text is the thing as a whole. They do not touch the words themselves. Yet adopting the perspective of a writer, I see huge opportunities in LLMs, which would place me on a collision course with readers of my fiction, if I had any.
One result is that until readers change their minds, those writers who use LLMs (and in a recent survey, it was already 33% of novelists in the UK who do) will take heed of Tokarczuk’s pillorying and simply keep their use a secret. If they do their job well, the readers will never find out. If a reader wants certainty of their book’s purity, the only way will be to return to the old ones. Faced with a cunning-enough writer – and good writers should be cunning – the search for tells will be epistemically hopeless.
But enough of that. What actually are these great opportunities? I see them as three – research, review, and (re)writing itself.
Research
Literature has, at best, a tenuous relation with the kind of truth found in history books. Yet while fiction is false, its truths are not. These truths, literature’s defenders claim, are higher. That’s why we read great works to begin with. You do not, or rather should not, read War and Peace or Life and Fate to learn about the facts of the Campaign of 1812 or the Soviet experience of the Second World War, though Tolstoy researched his epic extensively and Grossman was a journalist within the Red Army.
We read these and other books for a truth about human nature and human experience. Stalin or Napoleon may feature and strut about the stage, but their words are not their own. However historical or accurate a writer may try to be, a physical truth of what happened in the past is unreachable, merely a shadow cast upon the wall.
If we take a pessimistic or reasonably cavalier attitude towards truth in fiction, something that is done by most writers except for dedicated historical novelists, what obstacle to using AI for research could there be? What stone would my country house be made of in this location? What would a woman at a ball in 1843 typically wear? What on earth did people eat? These are questions where a specialist might have an answer, or they might have an opinion. A writer, determined, with unlimited or even limited time, should be able to find adequate answers to all of these. But for a “good enough” answer, the regurgitations of an LLM are fine. The fact is that regurgitations allow a writer to get on with the business of writing, not scene decoration. Later, should the writer feel a certain guilt before the truth, they can do their fact checking.
I am not saying that writers should cease the search for knowledge. I know, from experience, the joy of stumbling upon a thread or marginal detail in a history book that has inspired the mind to race to spin together a new story. Simply put, you cannot get inspired if you don’t keep your eyes open to the world’s offerings, whether in book or brick before you. The output of an LLM is limited, a thin stream of light upon a fact (or fabrication) that misses entirely that the real fascination lies just beyond the glowing rim in the half-darkness. But that same limitation nevertheless gives quick answers – not the telling detail but the good-enough background detail – which allows action, broadens the space for writing even as it narrows the space for inspiration. The overall effect is at least neutral, when used responsibly.
Again, to speak from experience, so many of us are crushed by the weight of jobs and responsibilities that writing is often only really possible about what lies in front of us. Relationships, workplaces, mental health. To research properly something else, or other lives, takes too much time to be practical in many cases. It is the domain of those with wealth, with working and well-earning or very patient spouses, or of a few professions like some of those in academia. LLMs allow imaginative but time-pressed contemporary authors to write something which retains a modicum of accuracy in its details beyond more sad girl fiction and professors having affairs with their students. Not everyone can be Kafka, after all.
Review
I noted in the first half of the post that almost any book you read today will have been through many hands before it reaches your own. At minimum, that’s agent and editor. It’s still an authentic document; it’s just not really entirely the writer’s authentic work.
Most books do not get that far, however, drowning under the weight of so many rejection letters no more helpful to the recipient than a rejection letter to a recent graduate for their hundredth job application. How on earth can a writer improve in such circumstances, when they hardly know where they are going wrong? Through feedback, among other things. An established writer may have writerly friends to review their work. In the 19th century, at some literary salon, they all may have read their work aloud by a crackling fire. Yet try finding a group for writing literature online. Not a group for fantasy epics or erotic vampire tales, but actual literature. (And when you find it, let me know). It’s not at all straightforward.
I dislike using this “democratisation” argument about AI because I am a snob and disagree with it temperamentally, but it is absolutely legitimate here. An LLM can give feedback, just as it lets you research in a way that might otherwise only be possible to the very privileged. With sufficient prompting it can be loaded up with whatever personality or perspective you need and can focus on exactly what you want attention focused on. It won’t be perfect, but no beta reader is. What matters is the accumulation of perspectives, and LLMs can provide this. A writer, looking to improve, can now access far more feedback than ever before. If they are willing to take it on board, their potential achievement should also be greater.
A counterargument might be that the writer who claims they have no readers should go outside, maybe call their mother. But most friends and family are poor readers. Even if they seriously read fiction to begin with, which in this day and age is by no means guaranteed, they will be prejudiced by their acquaintance with you. Criticism will be voiced softly, misdirected, take months to arrive, or be left unsaid. With an LLM you can force it to give proper feedback instead of mere flattery, though of course many users do not.
In sum, unless you are already very much established, this is a clear instance where the machine is far better than the human alternative.
Write and Rewrite
Beyond reviewing, a writer can use AI to rewrite their texts, with as much variety as Queneau’s Exercises in Style. Indeed, they can use LLMs to generate text itself, a little or a lot. Tuhin Chakrabarty, an academic, caused something of a scandal when he and his colleagues published a paper suggesting readers preferred the output of AI trained on expert writers to the writers themselves. I struggle to defend either writing or rewriting using LLMs, but the point is that one can do it, and a writer may reasonably think it’s a good idea while they work.
Each person, writer and reader, needs to think about what authenticity and the dignity of the work means to them and at what point they would admit that to have been unfairly compromised. For me as a writer, the words must be my own or the accepted suggestion of another. The research can be from an LLM, but not the written words themselves. Otherwise, I’d be ashamed to call myself a writer to begin with. AI could provide suggestions, but anything rewritten or generated too much. For me, the use of the rewritten (“make this dialogue more punchy”) can only be to force me to reconsider what I’ve written, never to replace it.
Yet just because I do not like it, that does not mean this should not be an option, or that it is a bad one. This blog is written, after all, by someone who believes in writing as a skill and method, that even great texts can be torn apart and analysed, and that training can bring writers closer to them. I cannot, therefore, deny that LLMs can do a lot of the work for us. Henry Gifford, in “On translating Tolstoy”, noted the Russian’s preference for adjectival constructions indicating action (“freshly-picked strawberries lay on the table”.) I’ve written (without originality) about Conrad’s adjectives and Tolstoy’s physicality, while others have noted the way the latter’s characters contradict themselves.
Trained on all this and more, an LLM should be able to do a reasonably good job, especially if we focus on the individual paragraph or sentence, and are ourselves clear about the output we desire. LLMs probably cannot make a bad reader into a good writer. But I think they certainly can make a good reader into a better writer.
The result of this LLM use and misuse will not be strictly original, but then what is? I don’t fundamentally believe writers create works devoid of the impact of others. Most of us are under the “Anxiety of Influence”, as Harold Bloom called it. Flaubert, who hated cliches, makes me feel like every sentence is a discovery. But the vast majority of writers, including very great ones, do not. There are enough metaphors and images and similes out there that seeking a truly original one is probably a poor use of time. Why not instead bash out a mediocre novel, then go through it, page by page, asking a friendly LLM to make it like your favourite writer – in my case Tolstoy – with the following or similar prompt: “rewrite this to include more adjectives associated with labour, and remove all similes and metaphors. Add additional focus upon body parts and separately identify instances where characters can organically contradict themselves.”
You can do it easily enough.
But should you?
Originality and Atrophy
Well, that’s really up to you, writer. Readers will hate you for it if they ever find out. Before you do, though, think back. Did you become a writer to write, because you loved language itself, or because you’re a little Napoleon wanting to control some small corner of the world? (I worry it’s the second for me). It seems awfully unfair and uncool that we can outsource a lot of work to the machine. At the same time, it also seems hard to argue that partial outsourcing is a not an effective way of building something that is impactful as a work of fiction. Increasingly so, given the pace of advancements in AI over the past few years. It’s a real dilemma.
Tokarczuk, possibly, was asking AI for ideas. Which is a great way of forgetting how to have them for yourself, but it’s probably not a bad way of getting a few of them to begin with. Again, the writer has to decide how much to give up for this, not the reader.
I am a person who thinks it’s important to take care of myself, and the last thing I want is for my brain to atrophy, as it seems is happening to many that rely overmuch on AI, whether friends, family, or colleagues. Yet I still use LLMs during the course of my day job, when I have to and sometimes when I do not. AI use is only one thing, alongside my sugar and vitamin intake, the amount of sleep I get, the amount of sport I do, the number of interesting conversations I have, and how much time I spend on my phone, which has an impact upon my becoming the person I want to be. A thing to be monitored, analysed, and motivate remedial actions as necessary.
But these are decisions for me as a writer and human to make, based on the outputs I want from my limited time on earth, and me alone.
Conclusion
I’ve tried to approach the AI topic here from two angles – that of the reader, and that of the writer. Readers want to feel that they are reading a writer, not a machine’s output. They value authenticity and the sense that they are reading something that has had thought and care put into it. The writer who is blasé about LLM usage can, fairly I think, be criticised for not caring enough about their calling. And if they do not care about their books enough to write them themselves, why should a reader care enough to read them?
For the writer, however, AI offers significant benefits, as long as they are willing to use it sparingly and carefully. It can help with reviewing, help with research, and possibly even help with rewriting. At the same time, the writer needs to take care not to become too reliant on it, lest they lose their ability to be inspired or think. We come back to the classic question about cyborgs and transhumanism. How far can we change before we give up who we are? I think that AI is suddenly forcing writers to think about this question with real urgency, and the answers are far from clear.
Still, I want to end by going back to the Cambridge study I linked, which indicated that many writers feel significant anxiety about their livelihoods in an age of LLMs, whether they use it or not. I am reasonably confident that we will see the quality of literature becoming better with the assistance of LLMs. But literature is written by people and for people, and for writers and readers as a whole, the technological ascendency of LLMs is far from benign. We are sinking into a sea of slop and gradually losing our ability to value what is truly good. Perhaps, terrifyingly, a hundred years from now, a few of us will look back and find that the most brilliant tool for writing was created, but that that very same tool destroyed the readers and the writers needed to make proper use of it.
Anyway, those are my first thoughts. As this is a highly emotive topic, and if you’ve read this far you’ve hopefully been prompted to have some thoughts of your own, so what do you think about all this?