Fiction & AI - He Do The Large Language Model In Different Voices

The last time that I blogged about AI and fiction, I finished by suggesting that the most likely manifestation of AI in the future novel meant a return to some of the characteristics of literary modernism – polyphonic voice, pastiche, fragmented experience, irony, deep interiority. 

That I’ve not yet further explored that idea’s largely down to the fear that the pace of AI development being such, anything I say will be written with one foot in the clouds of speculation the other in the dustbin of history and consigned to it headfirst before I’ve pressed ‘publish’. 

This fear has not restrained AI-oriented futurists, mostly fixated on the idea of AI as a mimic of capitalist production, defense posture & global hegemony – all the wet dreams of the oligarch class, embodied in an affectless piece of software trained to be as rapaciously moronic as its masters.  So let’s deal with capitalism first – at least as far as it orchestrates the means of production in the publishing sideshow of the global entertainment spectacle.  I’ll come to the fate of the primary producer of fiction – e.g. me – later.

It’s now perfectly possible to AI-source the copy-editing, book design, cover art and – for all I know the marketing campaigns too – of any given piece of fiction.  No doubt the high-recognition-factor celebs who appear as purely nominal authors on the covers of their works are in the process of themselves being replaced by AI-sourced celebs, designed and generated to target just the right sub-demographics (assuming there’s anyone left in those groupuscules with the attention span to actually read anything at all).  But what about, you know, plot and, like, character? And the actual hard graft of writing all those words – and then getting them in the right order?  Publishers the infosphere over must want to AI-source all that too, freeing themselves forever from dependence upon those pesky deadline-missing, importunate, messy, bibulous, publicity-averse wasters-of-midnight-oil actual writers.  You know, the people that we’re all supposed to be here for in the first place. 

A  modestly controversial proposal: I don’t believe that you can AI-source fiction – at least not worthwhile fiction, as opposed to derivative, performative-spectacular mimesis of genre (which may well be what a lot of readers are going to be content with).  And its not because AI is not yet powerful enough or well-trained enough to do so convincingly or because it hasn’t ingested enough source material or because its human implementors have simply not given the AI a sufficiently detailed  prompt from which to work.

No. It can’t be done because it is an error of categorization, rooted in our failure to understand what is really meant by the ‘artificial’ bit of AI, and to understand what it is that writers actually do to with by and from words. 

Let me explain.

Top bard Tom Eliot gave us a clue about this over a century ago in his now suddenly prophetic essay Tradition & The Individual Talent. 

AI – what tradition? AI – what talent?  Eliot proposes that each new exponent of a form is influenced by and in turn develops and extends the body of work, or tradition, in which they create.

Let’s take the tradition element first:

AIs can be taught, by ingesting texts (usually stolen goods) of a literary form, and can when prompted produce a more-or-less competent – depending upon the care and precision with which the prompt is devised – pastiche of that form.  We’ve all played this parlour game.  Some while ago I prompted an AI to write a poem on the doctrine of atonement in the style of the metaphysical poets. The result was a technically competent, perfectly dull pastiche that wouldn’t have disgraced a bright A-Level student. By which I mean someone who understood a little about Christian theology, the rules of rhyme and scansion, but had never experienced the reality and experience of suffering. (I’ll come back to the implications of this below). What it didn’t do was look sound or deliver meaning anything like a genuine piece of 17^th century religious poetry really looks sounds and delivers.  The AI may have consumed the tradition but it had not in any sense occupied its boundaries, let alone extended or transformed them.  As a play in the imitation game it was redundant. 

So what can AI actually do within a given tradition?  It can consume existing products and mimic them – more or less poorly.  It’s ability to improve its mimicry is, I suppose, limited by two factors –

(i)                  the ingenuity of the AI’s implementors, and

(ii)                 the availability of original material for theft/ingestion. (i.e. the corpus of the tradition to be mimicked).

Can either be significantly extended? Attentive readers may remember that a couple of years ago there was a brief panic when it was realized that AIs-in-training were running out of human-authored content to steal, erm, sorry – ingest.  The obvious solution – get the AIs to produce their own bloody training content – ran into a very serious problem: AIs trained on AI-generated content themselves produced outputs of ever-diminishing quality (however you define ‘quality’ – accuracy, coherence, utility, interest).

How to move beyond this boundary?  Let us suppose that a hegemony of publishers implements a population of AIs that are trained on the existing content of a tradition e.g. the literary novel. The first generation of AIs are then prompted to produce their own extensions of this tradition. A second generation is then implemented and trained upon the content of the first – they then produce their own generation’s content. A third generation can then be trained with the content of the second. And so on. Will innovations of form, voice or content arise? Will schools or movements detectably develop and differentiate themselves?  At what point will the horizon of commercial viability, or even reader readability, be encountered?  Will some kind of identifiable AI aesthetic arise?

My guess is that sooner rather than later for some genres, but that it will be infinitely postponed if the content / quality squeeze alluded to above holds good for most forms of production.  And the clue for the reason is in the name – we’re dealing here not just with artificial intelligence but with artificial creativity (and yes, in the real world these two things don’t have a quantifiable relationship either).

Now let’s deal with the other variable in Eliot’s calculus – individual talent. 

An AI can produce fiction that conforms to a given specification in mimicry of a given pre-existing tradition. But it cannot undergo the individual experience that a human being mines and transforms when they extend that tradition. Whether its family romance, exile, heartbreak, war, poverty, illness or thwarted love, whatever it is that feeds the writer besides the awareness of what has gone before – and I’m not going to be so jejune as to suggest that this must be solely a history of trauma or suffering – none of this can (currently) be experienced by an AI. Eliot gives us a clue about this problem of individual sensibility in his other great piece of criticism The Metaphysical Poets with his remark about the lost unity of intellect and sensibility, exemplified by how we experience the scent of a rose.  No doubt some of the tech broligarchy – if they can justify taking time off from the design of engines mimicking the more ruthlessly inhuman forms of capitalism - are today hooking up a spectrograph to a neural network, putting it in a garden of Hybrid Teas, and teaching it to smell the roses.  Will this mean that – as per Eliot’s dictum – it experiences the scent of a rose or simply understands it? I’d hazard that even with its integration into advanced robotics AI is still necessarily and unavoidably artificial intelligence, not artificial sensibility (still less the real versions of either) and that AI-generated poetry or fiction in fact represents a cultural end-state for the dissociation of intellect from sensibility that Eliot detected as beginning in the 17^th Century and which still characterises our experience today. 

So what does this technological and cultural squeeze mean for future fiction, assuming that any of us are still around to write or read it?  And how can fiction represent individual consciousness and social experience in a world of AI?

The tropes of modernism are a good starting point for this - interiority, multiplicity of voice and consciousness, innovation of language and form, fragmented experience and memory, representation of the sub-conscious. (I’d conjecture that it will prove impossible for an AI to have one).   

To take the example of two (comparatively recent) masters of late modernist fiction – Lawrence  Durrell and Alasdair Gray.  Both write complex, many-layered and polyphonic novels that are ‘about’ AI – though neither of them uses the term – in the sense that they feature intelligent, learning creatures who raise questions about consciousness, language, memory, the social self, etc. 

In Durrell’s 2-decker novel The Revolt Of Aphrodite a world-dominating plutocrat (who does that remind us of?) uses robotics & a form of AI to create a simulacrum of a dead screen-goddess with whom he had been obsessed before she was famous.

Alasdair Gray’s Poor Things uses a Chinese box of encapsulated narratives to tell the story of the implantation of a baby’s brain into the body of an adult woman revived from death. This triggers a chaotic and frantic learning process of what it is to be human (& particularly, this being Gray, sexual).

(Culture war alert: In both cases, male writers have imagined male creators (one a doctor, the others the plutocrat & his tech-bro narrator) creating female intelligences for whom experience is primarily imagined to be sexual.  In both cases the creations move rapidly beyond their control, wreaking havoc in the world.  Where have we encountered this before?) 

Both Durrell & Gray excel at the long slow burn of sliding-doors revelatory irony.  And its notable that jokes are one thing AI doesn’t seem to do well.  It could be that irony – the deepest joke of all - requires both writer and reader to be able to acknowledge and evaluate the authenticity of the experience of another voice, and the reality of the consciousness which produces that voice. The word is, I suppose, imagination. And it cannot be generated solely by consuming and mimicking the imaginative productions of others. 

Another conjecture:  as the transmitter of inauthentic experience and unearned emotion, the characteristic voice of AI-authored fiction will be essentially sentimental.  This will allow it to fit into a capacious niche within the prevailing tastes of American culture. 

Some end-thoughts:  to suggest that fiction should occupy only the republics of experience in which AI falls short would be to sell short fiction itself.   The assertion probably depends upon whether you see the novel as a distinct cultural form developing out of early modern and Enlightenment culture and society; or if you see the production of fiction – tale-telling, fabrication, deception, entertainment, the representation of social reality and the individual consciousness’ engagement with it, as a manifestation of 3 million years of hominid evolution in ever more complex social groups.   In such a tradition, each novel is an end-point of creation, a boundary of the universe – or at least the linguistically and socially perceptible boundary of that universe. But I would say that, wouldn’t I?  I’m a novelist. Reject all imitations.