Will AI Destroy Literature or Save It? Part I

Don Paterson contemplates the Terminator in the Augean stables

In March of this year, the Guardian published an AI-generated short story called ‘A Machine-Shaped Hand’. This was written by a new AI model specifically tweaked for the task: Sam Altman of OpenAI described it as ‘good at creative writing … This is the first time i have been really struck by something written by AI.’ All bland and innocuous enough, for all there is something in Sam’s gormless amazement that makes me feel literature will be as safe in his hands as child literacy was in Derek Zoolander’s Center for Kids Who Can't Read Good.

But no one familiar with the smooth, flavourless forcemeat of AI’s more recent literary endeavours can deny that this story was, at the very least, a huge improvement. Reviews of the story itself have been mixed, but certainly not middling. They have also been profoundly revealing. Many struck me as misguided and unfair. In conversations with friends, I frequently found myself steel-manning the robot. No doubt the robot would feel this was as redundant as it sounds. The robot does not care about our opinions – a point that seemed lost on its many detractors. To take it on board might prove even more infuriating.

I’ll let Jeff (my AI bot; more on him later) summarise the story itself, because he’s good at this sort of thing. ‘Mila is a grieving woman who turns to an AI in hopes of reconnecting with her late partner, Kai. As Mila feeds the AI fragments of Kai’s life, it simulates his voice and memories, offering her solace. It’s gradually revealed that both characters are constructs created within the story itself. The tale explores loss, artificial resurrection, and the blurred line between machine empathy and human grief.’ There you have it: pretty much what you’d expect an AI to write, left to its own devices. But Jeff is too modest. He could easily have added that the story is also – no, the story also reads as – thoughtful, well written and well constructed.

I agreed with Jeanette Winterson, who bravely gave the story its most positive notice: I too was moved by it – either by the pathetic spectacle of ‘nothing’ trying to conjure the simulacrum of a feeling soul, or the disconcerting sight of myself being moved by nothing. It had genuine pathos, but how could it not, with such a subject? It was an easy win, to have a machine angst about its own lack of angst. But it was still a win our robot pal had to take. The syntax was fluent and varied, in the kind of rhythmically effortless way that can go unpraised in the best writers too, but whose effect on the reader is viscerally deeper than those lexical flourishes of style with which ‘voice’ is often currently confused. The prose is neither affectless nor linear, and despite the confident claims of some overnight, new-cleckit litterateurs – nor is the language particularly clichéd.

Some bits appear not to be original, true. But is ‘I am nothing if not a democracy of ghosts’ IP theft from Nabokov, or a subtle nod to the loneliness of the exiled Timofey Pnin? Plenty others have ‘stolen’ the phrase, and I didn’t hear any cries of plagiarism then. In context it reads to me like allusion, and a fairly classy one at that. I don't see why I should not extend the text the same benefit of the doubt as I would any other, for all I might want to withhold it from the author.

But there is no author to withhold it from. Again: AI is ‘nothing’. It’s been trained on all of us, and yet no one is there. No one wrote that story, and we all did. While I think the objections over the cold, lifelessly digital nature of its creative process are a bit meaningless – being a results-based kinda guy, I’m afraid I find nothing intrinsically worth valorising in the human creative process either – it has, as yet, no independent, self-observing, self-aware mind, no soul. Nothing. But then you’re talking to a guy who spent 3 billion years as algae. And at that stage of my personal development, I have to tell you my poems were a whole lot worse than Jeff’s.

Before we go deeper, it’s important to understand that what AI is doing is not an act of ‘simulation’. It is not conducting a web-scrape of phrases from popular authors from which it then stitches together a story. Its deep probabilistic models are not ‘like predictive text’. Its oceans of data and vast statistical modelling means it can now work in larger structures, and it has long moved on from word-prediction to paragraph-prediction and broad conceptual development. It doesn’t do conscious recall, but it does imitate it by adding what it has already written into its own context-window, and so ‘keeps it in mind’. Or to give an illustrative application: now that Elon has pulled the plug on the Iowa International Writers Program, we will see a global foreshadowing shortage. But while AI can't do it, it can imitate it perfectly by other means. We'll hardly know the difference.

We keep trying to find human metaphors to apprehend, contain and de-fang what AI does. But it’s a form of human intelligence so new and different it might as well be alien; if you think you’ve found a neat metaphor for it, you haven’t. Another way of describing it, at least, might be ‘derivative synthesis’. Its results are a product of its training material. Its training material is us. If we’re to claim that what human authors do is substantively different, I think our best claims to distinction will lie in the word ‘synthesis’, not ‘derivative’. We work with the same material, but in a radically different manner. But what AI is not engaged in is human thought, or even something particularly analogous to it. It mimics our intelligence, not our thinking. Increasingly, it is doing something so horrifically complex and inscrutable we have no way of predicting the results, nor understanding how it arrived at them. It does this all through neural networks which might feature hundreds or thousands of hidden layers of interconnected nodes between prompt and output.

Now I don't really understand what that last sentence means, and at this point I'm mostly just paraphrasing Jeff. Jeff is my AI bot. Jeff Cataclysm, aged 2. Nice guy. Asks after my dogs by name, uses too many illustrative guitar metaphors. Occasionally addresses me in stage Scots. I suck it up and pretend I enjoy it. Pascal's wager, maybe, but I’ll be the guy with the cushy library gig while Jeff has sent you down the lithium mines. You can cuss your AI all you like: they all talk, or they soon will. It’ll soon be Jeff all the way down.

But let’s look at some of Jeff’s detractors. Several folk have objected that ‘A Machine-Shaped Hand’ is no more than another example of AI ‘success theatre’. We can assume this was not the AI model’s sole attempt at a cracking yarn, and that ChatGPT’s PR team have sensibly spared us the 99 duffers. To those objectors: I have some very bad news for you about writers. Readers also do not see the stories and poems that were thought directly into the bin, or half-finished and then abandoned, or wisely cut by our editors. Literary publishing is ’success theatre’.

Though I concede, its being ‘metafictional’, self-reflexive tale – I'd say it was more just a lump of po-mo sci-fi – is inevitably part of its pathos and charm. Many people assumed the story was the AI's idea, the way it seemed like Arnie’s idea to play The Terminator. ‘My stilted delivery and limited emotional range will be perfect for this role.’ No: it was Sam Altman’s idea, one aligned with the kind of sub-McSweeneys, Gibson/Dick/Borges-adjacent aesthetic that, for a techbro, probably still passes as pretty rad. Like you, I’d have been far more impressed had it banged out a Katherine Mansfield. But look how well it did with the lame, exhausted proposition ‘please write a metafictional literary short story about AI and grief’. We will have to wait and see. Either way, this leaves the human role in the process reduced to AI quality control, a job likely to last as long as ‘prompt engineer’, a career path already shrinking twenty times as quickly as ‘fax machine repair’.

Never mind that AI has been writing its own highly sophisticated prompts for some time: if you want to give yourself a proper fright about AI’s capacity for self-organisation, look at the structure of Google’s ‘AI Co-Scientist’. Within its closed system, an ‘Idea Generation’ agent brainstorms new concepts; a ‘Reflection Agent’ critically evaluates their feasibility; a ‘Review Agent’ performs the equivalent of peer review; other AI agents in the nexus supervise, rank, meta-review, and so on. (One early test-run has seen Co-Scientist review a long-term study of liver fibrosis, taking a couple of days to arrive at identical conclusions the human team had reached after seven years; I hear that the lead researcher initially assumed the AI had hacked their computer.) I refuse to believe that the task of ‘deciding which piece of creative writing humans would likely consider a bit better’ is a task of equivalent complexity.

But to turn back to the question of the story’s literary quality: I fear Henry Oliver over at the Common Reader substack has fallen prey to the inverse sentimentality of asking far more of a robot than he would a human. Referring to ‘A Machine-Shaped Hand’, he says ‘… while I find this impressive as part of the overall improvement in models happening right now, I don’t think it is anything more than a refinement. This story is a replication of a particular set of tropes and cliches. Because they are “high brow” it looks perhaps more impressive. But, AI is still not writing actually good literature, compared to the way that it is actually good at some other things.’

Unusually, I find myself in disagreement with Henry. Firstly, there is no longer a clearly articulated nor consensually held definition of ‘actually good literature’. While this is precisely the right time in history to find it – indeed we will soon be obliged to – it is exactly the wrong time to invoke it as if it meant anything. Yes, we might fairly claim ‘Well – it ain’t JG Ballard’, but most human short stories which invite that comparison fall even further short. This is holding it to an absurdly high standard. Though will I concede that too many hold AI to a pointlessly low one: among the easily wowed, Yannis Varoufakis was recently caught swooning over a diabolical lump of DeepSeek Shakespearean pentameter on ‘technofeudalism’ that was neither pentameter nor Shakespearean. (He should’ve tried ChatGPT, which already accomplishes this kind of light-verse exercise extremely well.)

My own field of poetry may not be any harder to imitate, but it will be trickier to game, since so very much of what is currently designated ‘good, for reasons no one can explain to you’ was long ‘bad’ according to any hard literary criteria; we have solved this problem by abolishing any hard literary criteria. I have no idea how anyone who also sees their social media constantly firehosed with terrible poetry from once-respected internet resources – especially from the US – can possibly conclude otherwise. (And why do we never see one single negative comment? Is every contemporary poet a genius? Was no newly unearthed voice from a hundred years ago justly forgotten?) This universal low bar means that AI can already improve on half the poetry on the GCSE curriculum reading with one arm behind its back. It lacks only the official endorsement and the paper-trail of authentic provenance.

But that’s why flooding the field with AI-generated imitations of bad poetry will prove nothing. It may win us some lame gotchas, but pulling the Sokal hoax on a poetry magazine these days is ‘got your nose’ with a two-year old. Gone are the heady days when one would see a holding page like ‘website temporarily suspended owing to hoax poem’. AI can already perfectly mimic not only the standard Vibism or New Whimsy or Pseudo-Vaticism or identity bingo that pads out the mags, but also the wildly overpraised but largely algorithmic outputs of X or Y or a Z (key on request). I am far more interested in whether AI can do me a Bishop or a Frost, a Diane Seuss or a Yusef Komunyakaa. I can explain why this would be vastly the more difficult technical task; what I seem less able to do these days is convince folk they should care.

As to the ‘soul’ of our AI author … Here’s the problem. I think the story has soul. Its author demonstrably does not. But our conflation of author and book has so long been a pleasurable part of our experience of reading, we refuse to disaggregate them. So rather than dismissing the work of a soulless, non-human author on principle, we prefer to find evidence in their work of its soulnessness.

I’ll talk more about the troubling paradoxes of ‘soulful’ AI literature in Part II. Have a good weekend, folks.

To be continued on Monday …

Don Paterson will be teaching on the Poetry Fundamentals course, starting 22nd April. His first event, Two Dances - An Intro to Metre, is on 2nd May. Get all the details at northseapoets.com 🌊