I

Was the Death of the Author, announced by Roland Barthes in 1967, an ‘accident’ of thinking? Like, or unlike, the swipe of the laundry truck that killed the writer one morning in the Quartier Latin – écrasé, crushed – a brutal word, close to erase – shutting down the body of a singular man who, later in life, had fallen in love with his own human singularity?

‘What did you do, Ray? Aw, shit...’ – Dr Peter Venkman to Dr Raymond Stantz, Ghostbusters (1984)

Camera Lucida, Barthes’s most poignant work, had appeared in 1980. In 1981, Yves Bonnefoy, whose work exudes a grave vitality very different from Barthes’s, was elected to the Chair at the Collège de France left open by the latter’s death in 1980. I was first attracted to Bonnefoy’s work by a poem about a lizard on a wall. It opens:

The startled salamander freezes
And feigns death.
This is the first step of consciousness among the stones,

and it ends:

How I love that which gives itself to the stars by the inert
Mass of its whole body,
How I love that which awaits the hour of its victor
And holds its breath and clings to the ground.
(‘Place of the Salamander’, translation by Galway Kinnell)

Bonnefoy elaborated a philosophy of ‘presence’. He tracked the movement of consciousness as it tries and fails to get a grip on the world: each image a tombstone closing over what it claims to represent. Reading this poem, I, too, freeze – and give my body up to stars – and hold my breath. I can never see a lizard on a wall in a garden without recalling these lines.

Yet it’s Barthes, his predecessor, who grips my attention when I think about the advent of large language models (LLMs to those who know).

I notice that poignant derives from the French for to puncture or to prick. The earlier Barthes had decisively skewered the Romantic authorial voice with its source in a singular body: ‘[W]riting is the destruction of every voice, of every point of origin. Writing is that neutral, composite, oblique space where our subject slips away, the negative where all identity is lost, starting with the very identity of the body writing.’ – Roland Barthes, The Death of the Author.

But the later Barthes would write: ‘The photograph is literally an emanation of the referent. From a real body, which was there, proceed radiations which ultimately touch me, who am here...’ (Roland Barthes, Camera Lucida). I’m touched both by the insight and the feeling of having-been-touched that the author confides to us in this later passage.

Perhaps the Death of the Author hadn’t carried away on its plural tide every point of origin.

II

Barthes proposes to examine ‘the phenomenon of photography in its absolute novelty in world history’ (The Grain of the Voice). I feel strangely moved by Barthes’s perception of this ‘absolute novelty’: a threshold that is crossed only once and for all time.

The unique impression made by light waves rebounding from an object generates a singular set of reflections. The physical impression of a material existence in time (an impression from that moment on detached from the being it reflects) constitutes the ‘absolute novelty’ which Barthes sought to pierce with his insight, reflecting on the exquisite sensations that ‘certain photographs’ made in his body and mind.

Human written language also detaches itself from the time and the place in which its meanings were composed by a singular body – thereby sharing a trace or impression of those meanings beyond that time and place.

Language can detach itself from material sources in other ways, too. Many word-randomisers and writing generators have existed over time, making the authorial voice a process of selection and curation from external inputs, underwritten by chance. A predilection for the offerings of chance is evident in centos and was fantasised by Stéphane Mallarmé in his 1897 poem ‘Un Coup de Dés’ (A Throw of the Dice).

Large language models are statistical machines for generating writing that take these prior interests to another level. Both randomness and likelihood are built into the generation process. Models such as ChatGPT, BLOOM and LaMDA can generate texts in many genres, including summaries, captions, dialogue and verse. Engineers and marketers refer to poetry as one of the genres of creative expression to which large language models will ‘democratise’ access (perhaps as calculators have ‘democratised’ maths, by doing the calculations).

But is it poetry?

Google’s Verse by Verse engine is an example of that effort. The Google Research website gives anyone access to its large language model, specially trained on the works of ‘classical American poets’. It can generate lines according to specifications such as ‘quatrain’, ‘couplet’ or ‘free verse’. It can produce a consistent number of syllables per line, ranging from 4 to 12 (or else, a random number) and it can tailor its outputs (though imperfectly) to one of four rhyme schemes (in addition to free verse).

The human writer’s tasks are these. First, to choose the ‘muse’ whose style will be mimicked in the output. Second, to initiate the generation process by giving the first line, known as the ‘prompt’. Third, to choose the verse type. Fourth, to click the ‘generate’ button and choose a line from the offered set. And finally, to give the poem its title. To test the process, I composed an example using Emily Dickinson as ‘muse’. I asked the program for couplets with eight syllables. Every line below (except the first line and the title) is exactly as provided, in the order generated. The attribution at the end was suggested by the model. Bees came up often in the lines it offered.

The Bee, my Brother

Call to me, my darling brother
White as the snow, as gold as snow.
She bent behind a song to pass,
Every one she sealed alas!
She had a daily drunken bee,
Who falls in rare variety.
She felt as if she held that night
Through centuries to dark, to light.
She had a daily drunken bee,
Who falls in timid honesty.

Written by the user
Inspired by Emily Dickinson

Parodic air aside, you will notice that some of the lines play fast and loose with physical fact (‘as gold as snow’) and spatial relations (‘She bent behind a song’). But we shouldn’t be misled. The source of this freedom is not an understanding of metaphor or the flights of a keen poetic licence that, observing the world, has chosen to transform it. Large language models simply lack any grounding in our physical reality. They do not know that snow is white or that a song has no dimensions in physical space. In the verse they make, word and world can never meet.

III

The advent of photography disrupted the reign of painting, and the appearance of writing overturned the oral epic. Certain values and cultural attitudes were lost. But so much was given to us that we never had before. Photography created the possibility of spontaneously recording important events, as they unfold, for posterity. Without writing, we would have none of the novels, stories, plays and books of poetry that we possess today.

So why do I feel rattled?

Perhaps because the large language models outdo all previous methods for engaging chance in the process of writing. They erase the ‘body writing’ entirely as a ‘point of origin’ for the production of words. In the space of language models – exchanging ‘prompt’ for ‘thought’ in the closing line of Mallarmé’s darkly glittering poem – ‘[e]very prompt projects a throw of the dice’.

And I think: never before have so many words (billions) from so many writers (millions) been scraped from the internet and poured (unaccounted for, deidentified) into the mixing bowl of training data for a set of algorithms.

And never before – given the power of patterns made visible by massive data sets – have the stirrers had such an (un)canny statistical knowledge of the sequences of language, generated by exceptional mathematics and engineering (with a huge, but greatly overlooked, leg-up from linguistics).

So here we are: language as literal grist to the mill of commerce. And another revolution in human expressive technology. Other revolutions have turned out rather well. So is my angst just a personal fear? Would I care if these models hadn’t taken aim at poetry, troubling the future of creative expression?

Far more urgent needs do exist for this technology, especially in health (say, translating for refugees when interpreters are scarce, or summarising conversations with your doctor).

Instead, the Death of the Author has taken on a physical form, stamping fear into writers. It’s as if we couldn’t help but think it into being: the large language model as Stay-Puft Marshmallow Man.

IV

An obvious question might occur to the reader. Won’t writers, by editing the outputs of large language models, create compositions that will, despite their automated source, reflect the writer’s purposes, their experience, desires, just as much as before?

Perhaps. And yet, there’s more at stake.

I find Barthes’s trajectory as a writer from The Death of the Author to Camera Lucida exceptionally moving at a moment in history when, for the first time, the presence of a human body is no longer essential to generate language.

Maybe that’s it.

This technological break is so powerful as to force our admission that a threshold has been crossed. As it was with the invention of photography and the advent of writing.

That’s why I want to bear witness to this change. Before the feeling fades.

And no doubt, the astonishment of ‘that-has-been’ [in response to the Photograph] will also disappear. It has already disappeared: I am, I don’t know why, one of its last witnesses (a witness of the Inactual), and this book is its archaic trace. (Roland Barthes, Camera Lucida)

V

I can’t help but see a relationship between writing as an animating force, and Barthes’s sense of certain photographs as a happening, a piercing and a puncturing of and by the reality of death: ‘I wanted to explore it not as a question (a theme) but as a wound: I see, I feel, hence I notice, I observe and I think.’

(Life’s animation by death is paradoxically potent.)

This quote turns René Descartes’ famous declaration, ‘I think, therefore I am’, on its head. To be conscious is to feel. And feeling is the ground on which a thinking being stands: ‘the condition of an interiority which I believe is identified with my truth’ (Roland Barthes, Camera Lucida).

I repeat to myself: the ‘condition of an interiority which I believe is identified with my truth’. And I think (because I feel): What if human writing emerges from the singularity of a human body (sensory organs, hormones, nervous system, and the whole embodied mind) that has developed, like a photograph, from every impression it receives throughout its growth – all the accidents of life, good and bad – in a process that doesn’t end, for as long as those impressions continue to be made upon and held by the body?

And I think (because I feel): What if pleasure and pain are resonances, struck and amplified by each of our encounters with that – where that may be a word, an idea, a texture or a touch; a face or a name, a gesture or a voice – that which – astonishingly, for what are the odds? – is attuned to the complex resonator that developed inside my own singular body? And what if those ‘frequencies’ could be thought of as the unique signatures of my body in contact with this world – like the knots and the twists and the rings of a tree that has met with strong winds or the interference of other branches in the course of its growth – but also mild conditions that allowed the laying down of sturdy heartwood?

And I hear again: ‘The photograph is literally an emanation of the referent. From a real body, which was there, proceed radiations which ultimately touch me, who am here...’ (Roland Barthes, Camera Lucida).

Responding to this passage, Michael Moriarty writes: ‘what is peculiar to photography is the unbreakable coupling of the image to its referent in the real world’.

But the advent of digital manipulation and image generation has broken that ‘unbreakable’ coupling. This only serves to underline the sheer progress (fundamentally driven by commercial investments) human beings have collectively made towards cutting off language and image from the material earth, including our body, which was once the only place in which they grew.

Barthes’s words defined photography. But could they extend to writing? Human bodies, too, are marked by emanations from experience. How could our words and meanings touch others if not by mind- and body­­-piercing sounds, resounding in the grain of all we’ve lived?

Barthes distinguishes the source of linguistic impressions from the (then) adamantly physical source of photographs on the basis of what he calls certainty: ‘No writing can give me this certainty. It is the misfortune (but also perhaps the voluptuous pleasure) of language not to be able to authenticate itself.’ And he writes: ‘the elaborations of the text, whether fictional or poetic, [are] never credible down to the root’ (Camera Lucida; original emphasis).

Yet, reading back his own words, I wonder, did he not feel certainty? Did those words not bear witness to their source in his singular existence? Certain passages of Barthes’s touch me deeply with their gestures and their meanings. These passages speak to some formation in myself, and set it ringing with a strange new animation, however many times I read these words.

They feel credible down to the root.

VI

If human language was once imagined as the place where existence speaks through us, who or what speaks in generated writing? ‘The place of the real in all this is problematic’ as Michael Moriarty says of Barthes’s ‘Death of the Author’ phase.

I think: we have absented ourselves from our common creation, language, as never before in human history. And with that absence, we risk devaluing, even abandoning, the vital connection between language and the real.

VII

Like Bonnefoy in the salamander poem, I love that our consciousness can mediate our access to material reality. Stitching our being and the outside world together, in a rapid shuttling motion, every moment of our lives.

That’s why, if you tell me that large language models are writing poetry today, I will tell you: our definitions of poetry differ at the root. The root – contested as it is – that separates machine from human being, and grafted automaton from simple conscious life.

VIII

For Emily Dickinson, the process of writing (both herself and her poems) was imagined as a forge. That forge was lodged in the singular existence she referred to as a Soul. She asked a question that still feels audacious now: ‘Dare you see a Soul at the “White Heat”?’ The poem closes:

Least Village, boasts its Blacksmith
Whose Anvil’s even ring
Stands symbol for the finer Forge
That soundless tugs – within –

Refining these impatient Ores
With Hammer, and with Blaze
Until the Designated Light
Repudiate the Forge –

When I write, I test ideas, words and metaphors against my inner ‘anvil’ to hear how they ring (right or wrong, sharp or flat). Those ‘soundless tugs – within –’ are minute judgements that carry with them the subtle gradations and varying intensities of a positive, neutral, or negative feeling – a pleasure or revulsion. It’s these that prompt my choices. And at times, the tugs are so intense that it feels as if the brain catches fire with the flames of a metaphoric heaven or hell. Reading and writing can be a gift or gutting.

Dickinson felt this when reading others’ books. But I think the same happened when she read her own creations, testing each for its capacity to freeze or burn her soul: ‘If I read a book [and] it makes my whole body so cold no fire can ever warm me, I know that is poetry. If I feel physically as if the top of my head were taken off, I know that is poetry. These are the only way I know it. Is there any other way.’ Dickinson’s ‘finer Forge / That soundless tugs – within –’ is simulated by the statistical machinations of large language models, which, ranging over trillions of possible sequences of words, select those strings whose frequency of co-occurrence in the training data and aptness of context suggest they can match the semantics of a prompt. Statistical ‘weights’, then, are the triggers for the choices that they make.

As it stands, language models still require our input. But in future, even the initial instruction or prompt (say, ‘write a sonnet on the death of my mother’) could be a sequence of words generated by another language model. A third model may adjust and refine the written outputs. There needn’t be a human here at all.

And yet, I insist, it won’t be poetry unless a human being, with their inner forge blazing, is somehow, somewhere, present in the process – having learned with sweat and tears, at the forge of trial and error, ‘how to weigh the harvest of light’, as the late Australian poet Robert Adamson wrote.

IX

One way to read Emily Dickinson’s poems is by passing through the thicket of her variorum of drafts, documented in Sharon Cameron’s book Choosing Not Choosing, which follows the written traces of those ‘soundless tugs – within – / Refining these impatient / Ores’.

For the architects of language models, it’s easy to imagine the human writer as a decision-making machine – choosing / not choosing – and human intelligence as a complex, but replicable, sequence of perception, decision and action.

In this model, human writers might think of possible patterns or distillations of words or ideas (perceiving the contents of their mind). They might then take the time to choose between them, presenting the results of their choices to the reader.

In similar fashion (if you accept the premise), a large language model generates words, chooses the sequence with the strongest rating, and offers that sequence. For language models, writing is indeed nothing more than Barthes’s ‘neutral, composite, oblique space’.

But lacking a body, and without access to the physical world, language models can neither know nor feel the harm (or the good) their words could do.

X
Barthes wrote: ‘the photograph of the missing being, as Sontag says, will touch me like the delayed rays of a star’. Certain poems can do the same.

Barthes referred to his ‘“ontological” desire’ to understand photography and what this technology did that was entirely new to history. I feel the same desire in regard to language models.

In 1882, Nietzsche wrote: ‘Lightning and thunder require time; the light of the stars requires time; deeds, though done, still require time to be seen and heard. This deed [the death of God] is still more distant from them than the most distant stars – and yet they have done it themselves.’ That delay is why we need to bear witness, even as they unfold, to the impacts of large language models on consciousness, poetry and other language arts. Barthes wrote in Camera Lucida: ‘What matters to me is not the photograph’s “life” (a purely ideological notion) but the certainty that the photographed body touches me with its own rays and not with a superadded light.’

And so I come back to the idea of resonance – the strike of the hammer on our innermost anvil – and the adventure of the writer in their search for human company. The company of those who show us, through a detail or an insight, that we are not alone when we meet in certain poems, as Barthes did in certain photographs, ‘the wakening of intractable reality’.

References

Barthes, Roland, Camera Lucida: Reflections on Photography (New York: Hill and Wang, 1980)
Barthes, Roland, The Grain of the Voice (New York: Hill and Wang, 1985)
Bonnefoy, Yves, On the Motion and Immobility of Douve. Translated by Galway Kinnell, with an introduction by Timothy Mathews (Newcastle upon Tyne: Bloodaxe Books, 1992)
Cameron, Sharon, Choosing Not Choosing: Dickinson’s Fascicles (Chicago and London: University of Chicago Press, 1992)
Google Research. https://sites.research.google/versebyverse/. Accessed 12 January 2023
Moriarty, Michael, Roland Barthes (Cambridge: Polity Press, 1991)
Nietzsche, Friedrich, The Gay Science, trans. Walter Kaufmann (New York: Vintage Press, 1974)