Closed Room 1: Experiments in Lockdown

Conflicted Copy was composed during a period of extraordinary global constraint. As his prefatory note records, Riviere experimented with GPT-2 between December 2020 and January 2021 when England and Wales were in a third lockdown. In that period alone, 47,855 people in the UK died from the virus and a further 1.5 million tested positive. The Covid 19 pandemic is a minor presence in the collection, occasionally registering in biting references to a future ‘Where the most / difficult decisions are finally put into the / hands of real people’ and dreams of safety ‘in an era that I cannot see beyond.’ (‘Safe PDF’); or in ‘Safe Dogs,’ where ‘the only thought that enters your head is “The / world must be falling apart around me!”’¹³ The virus’s submerged presence in the collection reflects anxieties that AI technologies, which corporations like NVIDIA, Apple, Microsoft, and Alphabet have insinuated into our social fabric with astonishing speed , will foster isolation and division.¹⁴ But it registers, too, in the way language is configured.

LLMs use and understand language based on the distributional hypothesis developed by Zellig Harris in 1954. Harris argued that words that occur together are more likely to share meaning. Thus, where ‘tree’ occurs, ‘leaf’, ‘bark,’ ‘root’, and so on will frequently appear with it; or, in the case of familiar phrases, ‘at the end’ and ‘of the day’. The distributional hypothesis fell out of favour among linguists in the 1960s but was adopted by researchers in natural language processing, who realised that a machine may be taught to understand distributional semantics by ‘embedding’ words as tokens or vectors in a high-dimensional space. The machine could then be instructed to predict which word will follow the last word in a sentence based on its proximity to other tokens.¹⁵ The LLM’s preoccupation with probability and distribution uncomfortably mirrors the fact that calculating the probability that proximity would lead to transmission was a matter of life or death during the pandemic.

Riviere’s process of composition exploited the LLM’s capacity for mutation. Learning how to prompt effectively involved figuring out how to break the most insistent lines of transmission between words:

Quite quickly it [became] clear that it would want to go down fairly well-worn paths, […] so if you put ‘at the end’, it’s going to suggest, ‘of the day.’ It’s attracted to the well-worn grooves of language. […] So, the negotiation became about how to get it out of those troughs and push it into areas where it would make weirder […] suggestions.¹⁶

In procedural terms, Riviere’s efforts to coax GPT away from language’s ‘well-worn paths’ are a kind of detournement: a radical diversion of language to reveal or restore its latent subversive dimension (everyday behaviours and gestures can also be détourned, as evidenced by the social distancing practices that emerged during the pandemic). As he became more accomplished, the poems evolved into stranger and more complex compositions; becoming, as ‘Safe Poem’ puts it, “less static, more intricate with age.”¹⁷ The collection is arranged in the order in which it was composed, tracing an arc that also illustrates Riviere’s developing relationship with GPT-2: the earliest poems are shorter and simpler, like experiments in prompting a coherent response, with a tone that is often blandly reminiscent of instruction manuals. As the collection progresses, however, they become darker and eerier, conjuring disturbing scenes such as the haunted ‘Pink Mode’ in which ‘the space beneath us’ is inhabited by ‘unintelligent’ creatures, or the ‘body snatchers’-like scenario in ‘Old Dogs,’ where nameless figures are reduced to ‘empty shells’ whose ‘lips moved, but no words / came out’.¹⁸

The sense that ‘We live in the / present and its image is not ours’ (‘Pink Colours’) is felt throughout the collection in doppelgänger motifs such as in ‘After Colours’ (‘an observer whose / existence is a reflection of that other presence’).¹⁹ Both the first and second poems in the collection (‘True PDF’ and ‘True Souls’) trouble the stability of the poetic voice. In the former, ‘a / gentle voice, singing as it speaks,’ suggests both an affinity with lyric tradition and that this tradition might be expanded to include a non-human voice (‘as it speaks’). In ‘True Souls,’ the speaking I appears to observe human behaviour at a remove (‘The human imagination can conjure wonderful concepts. / Sometimes I think they actually enjoy it’).²⁰ But the abiding sense of disquiet as to whether a human or algorithmic voice is speaking is never dispelled—as acknowledged by the speaker of ‘Darken Colours,’ who notes, ‘I found myself particularly distracted the whole time / reading both versions’.²¹ The poems appear to have a particular speaker both because Riviere selected words that took the poem in this direction, and because the LLM responded to his selection each time in calculating which would be the most likely word to follow. This process of mutual attunement conjures an uncanny and ultimately irresolvable hesitation in the mind of the reader, as to whose voice they are reading.

If Riviere’s I and you have a viral-like tendency to double, this is simply a reflection of material fact: an LLM is not singular, but a distributed system shared across multiple data centres. As Katherine Elkins states, ‘machines are polyvocal rather than univocal, containing multitudes rather than a single author.’²² In any given poem in Conflicted Copy, answers to Riviere’s first prompt may have come from a data centre in northern Virginia; the second from Newport in Wales; the third from Hohhot, China; and so on, crossing the globe at extraordinary speed. Where this distributed self appears to cohere is in the enclosed space of the dialogue box—another kind of closed room, in which the notion of a cohesive selfhood becomes difficult to sustain.

What mediated this negotiation with a vastly distributed network, according to Riviere, was language. In his account of his process, he suggests that language is the real agent behind the collection: ‘Language itself produces this,’ he informed the Words That Burn podcast; these were ‘thoughts the language allowed me to have.’²³ This is language presented both as oracle, ‘speaking itself’ in the manner of Bernstein, or Cage’s use of the I Ching, and language as a form of enclosure (the disciplining connotations of ‘allowed me to have’). For instance, see ‘True Mode’:

Now you’re done with

the usual paper and ink

(unless this sounds like a

bad idea), you’re free to

wander out of the study

room into the open air

and see what’s waiting

for you: a view of some

distant city with its

towers looming on the

horizon; some more

mountains to your right;

a vast river flowing

below you in the same

way you just read about

in your notebook.²⁴

The poem has a chiasmic structure, bookended by references to texts or writing. If the journey from the enclosure of the study out into the world recalls lockdown’s sense of ‘outside’ as provisional and restricted, the end of the poem suggests a more fundamentally disembodied state of being, as if the outside depended upon language to make it real. The room so hemmed in by language that the real world is unreachable recalls the sealed chamber in Searle’s ‘Chinese room’ thought experiment. This is the world as the LLM experiences it: contained by language, and only accessible via language’s mediation. Language, in the form of the texts that comprise its dataset, is the limit of its reality.

Closed Room 2: The Dialogue Box

In Riviere’s process of prompting and selecting words generated by the LLM, the equivalent of Searle’s closed room is the white rectangle of the dialogue box, a kind of oracular space of limit in which poet and algorithm meet one another in mutual incomprehension. Here, the I as the focal point of conventional lyric poetry becomes the cursor—not the icon of a singular consciousness, a point of focus, stability, and cohesion, but the sign of a radically distributed presence. On the one hand, this is in keeping with Riviere’s scepticism about the claims and affordances of conventional lyric (“Poets should recognise that the I in their poems is not conterminous with the legal entity who wields rights of ownership,” he states in an essay called ‘In Defense of Poetic Plagiarism,’ ‘but is fundamentally a fiction, a figment, a fold—a blinking cursor’).²⁵ But in the composition of Conflicted Copy, where both Riviere and GPT was represented in the dialogue box by the same flickering sign, the I-as-cursor becomes a point of rupture with the very idea of a cohesive selfhood.

Unlike traditional lyric, ‘digital work does not confirm selfhood,’ writes David Jhave Johnston. ‘The digital form operates as a container for aspects of a networked self that can only be acquired there.’²⁶ It isn’t only the LLM that is distributed between datacentres; traces of the poet (and every user of the web) exist scattered across globe, encoded as digital information, including some of the most intimate or cherished aspects of ourselves. As in Riviere’s previous collection, living online is a source of anxiety. The speaking voice of ‘Old Colours’ seems to treat the very notion of a cohesive, essential self—‘the part that is often called “the core”’—with suspicion.²⁷ Interiors frequently open out into uncanny dimensions, and several poems are concerned with journeys through ‘inner space’ into a kind of immateriality. ‘After Colours’ imagines a sci-fi scenario in which a team of scientists venture into an ‘unknown / realm’ that is also ‘inside their minds’ and where they find ‘strange creatures’; and ‘Darken Fame’ describes a journey ‘behind the curtain’ where it feels ‘as if something / sinister is going on.’ The poem conjures a Ballardian sense of inner and out reality folding in upon one another (‘Many who travel inside this / realm find they never completely step outside it’) wherein what lies outside the network is only available ‘through an increasingly blurred lens, / as they move from one layer to another.’²⁸ Moving through vector space becomes a kind of katabasis, descending further, layer by layer, into the architecture of the neural net.

Elsewhere, however, the cursor offers a different kind of transition. In ‘Safe PDF,’ it appears as a kind of aperture, a passage to the ‘closed room’ of vector space:

                    All that’s left is to save your

document by placing the mouse cursor

inside a closed room where only a trained

observer could look—when you put the

cursor on an invisible point, a small black

space will open above it, then another, and

if those spaces overlap (for the reader)

there’ll seem to be a tiny passage, in which

you can scribble any desired hole or curve

in space-time²⁹

Here, katabasis becomes anabasis, each layer in the neural net architecture opening upwards to greater and greater levels of possibility until we reach the promise of Artificial General Intelligence, the superintelligent machines that—so A.I. companies along with their investors and enthusiasts promise—will eliminate entire categories of work and usher in a utopian era of unlimited potential. This is vector space as dream-space, where the networked self is released from the closed room of embodiment (the poem opens in an atmosphere of surveillance, with an anxious warning not to ‘take off all your clothes / before getting into bed’) and into the manifold possibilities that vector spaces purports to offer.

‘Safe PDF’ is also a fantasy that occludes the very material reality of AI. As Ed Finn observes, behind the cursor icon, ‘you find billions of words authored by humans. You find thousands of hours of human attention and curation in the continuous “fine-tuning” of generative A.I.’, framed by ‘complex sociotechnical systems […] made up of corporate policies, shareholder commitments, sales ambitions, regulatory schemes, and internal memos’;³⁰ to which of course we should add the very physical data centres that house the systems and the harms they contribute to—the ecosystems transformed for their construction, the carbon they emit, and the water diverted to cool them; data annotators, millions of them in the Global South, forced to view thousands of violent images every day; open cast mines in D.R.C. and bomb impact craters in Gaza.

If direct references to the materiality of the AI industry are largely absent, other forms of AI-sponsored violence occur throughout Conflicted Copy. The role of artificial intelligence in political surveillance is addressed in ‘Safe Dogs’ (‘if the police are still in the area, you may risk / exposing yourself as an “enemy agent”’) and electoral manipulation in ‘Pink Souls’ (‘like a giant hand in a puppet show’). The way LLMs relay back to us the racism, misogyny and prejudice against disabled and trans people that are encoded in their datasets is acknowledged in ‘Darken Souls’, where the voice of the poem observes that what ‘sounds like / another’s words in your mind’ is in fact ‘a / product of your own darkness.’ ‘After Dogs’ gestures to the harms done by pornographic deepfakes and AI ‘girlfriends’ in an occult scene in which a young woman, her body exposed to a crowd, ‘whispered to the assembled audience that no Christian / male would touch or even look at her body whilst it could / be used for her own satisfaction or to please masturbators.’³¹ We might read ‘Darken Poem’ as A.I.’s mea culpa to those users who experience “serious emotional suffering and / damage” by believing their A.I. interlocutors are real, or a critique of the pathological need for self-deception that convinces us that LLMs reflect back to us a unified, speaking subject (‘the most obvious symptom of a / narcissistic personality disorder’).

As it progresses, however, ‘Darken Poem’ offers a pointed critique of the fallacy of immateriality on which the AI industry depends:

Darken the poem and it reveals the exact

reasons for this ‘deception’, not because it is

some sort of metaphor, but because the poet

himself has never been capable, despite

repeated efforts and promises, of actually

giving me the body I am dreaming of.³²

Darkening the poem reveals a fundamental failure to embody AI, that is, to realise or concretise the extractive reality behind the frictionless ideal. Jakko Kemper writes that ‘generative AI services are modelling a form of everyday existence in which aesthetics and ecological destruction become progressively interlinked. […T]hese services extend what Ulrich Brand and Markus Wissen have called “the imperial mode of living”—a form of existence in which quotidian comforts in one location are fuelled by extraction and exploitation elsewhere—to the realm of aesthetic productivity.’³³ A number of poems remark on the LLM’s disembodied condition, sometimes wistfully, as in ‘After Souls’: ‘I cannot describe in language how the sound of the wind and / waves become / like their reflection in ice’.³⁴ But while the machine may dream of a body and the world this would make available, ‘Darken Poem’ makes clear the poet’s obligation is to reveal the distributed but nonetheless profoundly material, world-impacting body it does in fact possess.

If the poet in ‘Darken Poem’ is said to fail in this regard, understanding how the collection as a whole meets this challenge depends on a re-evaluation of its limits. Kenneth Goldsmith describes the challenge of poetry in the digital age as one of ‘textual abundance.’³⁵ There’s no better illustration of this than the exponential expansion of the data pool on which each generation of GPT was trained. GPT-2 was trained on 8 million webpages, ten times as much data as was used to train GPT-1. GPT-3, released the following year, was 100 times larger than GPT-2, with 175 billion parameters; GPT-4 (released in 2023) was 1000 times larger again, able to navigate between 1.76 trillion parameters. As Lauren M. E. Goodlad and Matthew Stone note, the era of big data is also the era of data positivism, where the production of information at scale is enlisted by ‘a political economy that amplifies surveillance, the concentration of power, and the perpetuation of algorithmic discrimination.’³⁶ According to Conrad Steel, a ‘poetics of scale’ equipped to respond to this volume of information requires ‘a technique for bridging the gap between subjective experience and the massive systems of coordination and coercion through which that experience is organised.’³⁷ In Conflicted Copy, however, what we encounter instead is poetry at scale.

Much procedural poetry engages with scale and the prospect of infinite creation. Hélène Aji cites Mac Low’s ‘Willie’s Clatter,’ which draws on the Fibonacci series, and Ron Silliman’s Tjanting as works that reflect on exponential growth and limitation.³⁸ Goldsmith suggests that Darren Wershler and Bill Kennedy’s The Apostrophe Engine (in which hyperlinked list poems generate new poems whenever a link is clicked) ‘could be the largest poem ever written.’³⁹ None, however, approach the question of scale to the same extent as Riviere. The nature of its composition means that Conflicted Copy is not confined to the published poems. With each selection Riviere closed off all but one iteration of the emerging poem, but these possible versions still exist in data centres. Moreover, they exist as properties within the LLM’s probability field, and following each through its various iterations yields a truly astonishing number of possible poems. The shortest poem in the collection, ‘True PDF’, has thirty-one words; only the first word and the title were chosen by Riviere. Supposing that, in answer to each prompt, GPT offered 5 possible words, this creates a branching sequence of potential poems numbering 5 to the power of 30 (931,322,574,615,478,515,625, or nine hundred thirty-one quintillion, three hundred twenty-two quadrillion, five hundred seventy-four trillion, six hundred fifteen billion, four hundred seventy-eight million, five hundred fifteen thousand, six hundred twenty-five). Conflicted Copy therefore represents the tip of an unimaginably large heap. The greater poetic project of Conflicted Copy, of which the published collection is a tiny part, exists on a scale that exceeds the imagination—as close to infinite as any poet has come before.

Unlike other poet’s discarded drafts, the words offered by GPT-2 and rejected by Riviere remain a part of the wider project of Conflicted Copy because of their enduring presence recorded in data centres around the world. Raymond Queneau’s ‘A Hundred Thousand Billion Poems’ offers an instructive comparison. This vast number of potential poems are latent in the printed text, waiting to be realised by the adventurous reader who cuts out Queneau’s lines and remixes them. Their status as poems depends upon the combinatory possibilities suggested by the underlying procedure. But, as interchangeable printed lines, they nonetheless have a distinct materiality. Riviere’s prompts are likewise latent variations on the published collection, linked to it by the underlying combinatorial process, with a distinct materiality recorded in datacentres. Yet the scale achieved is dramatically different.‘A Hundred Thousand Billion Poems’ yields 10 to the power of 14 poems; ‘True PDF’ alone yields a number more than nine million times larger. Writing on this scale, Riviere’s poems have an unusually, even uniquely, material existence. Given that a single GPT search produces between two and three grams of carbon, it may be the first poetry collection whose carbon footprint can be calculated.⁴⁰ ‘True PDF’, for example, would have produced between 60 and 90 grams of carbon. The quantities are modest, but they underline how these are potential poems embodied in the world, on hard drives and in atmospheric emissions.

The words Riviere discarded aren’t realised poems, of course, nor are they accessible to the reader. But by persisting as tokens of relative probability retrieved from vector space and recorded in datacentres, they can be thought of as what Jhave calls ‘spoems,’ a neologism of poem and Bruce Sterling’s concept of ‘spimes.’ According to Sterling, spimes are emerging entities that are characteristic of a totally networked state, ‘material instantiations of an immaterial system.’⁴¹ Beginning and ending as data, spimes are profoundly relational—‘a set of relationships first and always, and an object now and then.’ Spoems, Jhave suggests, are ‘languaged spaces where poets establish networks of resonance between things and experiences, memories and intuitions.’⁴² The point of access to this material presence is, counterintuitively, the dialogue box, the icon of Searle’s Chinese room. When GPT’s five words materialise in the dialogue box they establish a node in a vast web of human connections, both residually present in the dataset and materially present and active in the world.

Closed Room 3: The O-Machine

LLMs cannot understand the world as humans do. For the machine, ‘the world’ consists of statistical constellations and meaning equals likelihood. Yet N. Katherine Hayles argues that even neural networks have a distinct umwelt or ‘world-horizon.’ Umwelten, a term coined by the German ecologist Jakob von Uexküll, are the discrete perceptual worlds inhabited by each species of animal. Uexküll’s famous example is the tick, whose world is defined by its sensitivity to temperature and butyric acid. Each living subject is at the centre of its own world, he writes, however constituted: ‘a simple world corresponds to a simple animal, well-articulated world to a complex one.’⁴³ As Hayles argues, umwelten are essentially information constructs, determined by the “architecture” (or physiognomy) of the subject. For an LLM, the world horizon is determined by the ‘the databases used in training them, the number and construction of the neural layers in their architecture, and other particularities of their algorithms and functioning.’⁴⁴ As noted above, language is the limit of an LLM’s reality. According to this view, they are very much engaged in making meaning when they interact with language, just of a radically different kind to that made by us.

Hayles argues that, like the profoundly ‘other’ intelligence of non-human species, generative AI’s distinctive form of intelligence can help us escape the confines of a positivist understanding of cognition. Technologist James Bridle observes that we have restricted our understanding of intelligence to that which looks like our own kind, restricting thinking to the closed room of ‘what happens inside our head’. Yet the possibility of what Bridle calls an ‘ecology of technology’ was recognised at the very inception of artificial intelligence. In his landmark 1950 paper, ‘Computing Machinery and Intelligence,’ Alan Turing proposed that the definition of a ‘creative mental act’ entailed an element of surprise, regardless of ‘whether the surprising event originates from a man, a book, a machine, or anything else.’ Turing understood that intelligence is an emergent property of interaction, Bridle argues, rather than a capacity locked away in the human brain: ‘intelligence doesn’t reside wholly inside the head or the machine, but somewhere in between—in the relationship between them.’⁴⁵ Intelligence takes much more than a human shape; in the midst of concerns around the disruptions and risks associated AI, from mass unemployment to the existential jeopardy heralded by Artificial General Intelligence, projects like Conflicted Copy can help us realise another kind of potential in our encounters with AI’s very different world-horizon: not just a radical expansion in our understanding of what intelligence looks like, but the realisation that intelligence is inherently collaborative.

Throughout nature, we find examples of intelligence as collaboration, made through relationship between beings, from the hive mind of bees and ants to Michael Levin’s insights into the embodied cognition evident even in individual cells.⁴⁶ From this perspective, a collaboration like Conflicted Copy represents a ‘cognitive assemblage’ of human and machine umwelten.⁴⁷ Hayles argues that accepting that an LLM can appreciate a sense of meaning, even only in terms of likelihood, is a step towards a more ecological understanding of cognition.

Riviere’s method doesn’t dispose entirely with what Perloff calls the poet’s “inventio”,⁴⁸ although it does subject this to a considerable degree of constraint. The question of which inventio—that of the poet or the LLM—hovers unresolved over every poem, line, and word in Conflicted Copy. But if we take seriously the agency of Riviere the poet, when that agency is restricted to choosing between half a dozen words from the entire word hoard, then we need also to consider the incalculable choices of those language users whose decisions have shaped GPT-2’s data set. Dimitri Coelho Mollo & Raphaël Millière argue that an LLM’s distributional approach to language is conducive to meaning, or what they call the ‘world-involving function’ necessary for ‘referential grounding,’ i.e. the capacity to connect words & phrases to real-world entities.⁴⁹ Crucially, the LLM establishes this via collaboration with humans, first via indirect causal contact in the dataset, and secondly in post-training, such as by fine tuning or via reinforcement learning through human feedback (RLHF); or when given ‘in context prompts’ such as Riviere provided when composing the poems in Conflicted Copy. To this, we ought to add a reminder that much of this ‘fine-tuning’ work is done by data annotators, many in the Global South, working in unregulated conditions for poor wages. The poet’s ‘I’ is subsumed among this vast matrix of language users, a tangled bank of inventiones. ‘Computers evolve,’ Hayles states, ‘through their partnerships with humans.’ But where AI is concerned, the ‘profound differences of embodiment’ have shaped ‘inverse evolutionary trajectories’.⁵⁰ Humans took millions of years to progress from environmental immersion to symbolic representation; computers worked with symbols from the beginning and have yet to achieve immersion. Several of Riviere’s poems examine this. ‘True Dogs’ reverses humanity’s movement from immersion in a world of animals to the animal as symbol so that it reflects the evolutionary direction of artificial intelligence (‘While we could continue this discussion / on the symbolism behind the use of / animal blood in these designs […] a / common and fundamental way of relating to humans seems to be that you / are just like a dog’),⁵¹ while ‘Dead Fame’ presents the evolution of human and animal relationships in terms that sound very like the training regimes of LLMs like GPT-2:

                                                                        Animals and

Humans have evolved from one another. They’re designed

with a very distinct relationship and understanding of each

other, a special type of mental communication between

beings that can only come with specific training. This

bond has been maintained through all the various scientific

advances.⁵²

It is striking that both poems are concerned with symbolism, one of the earliest and most powerful forms of human technology. The prevalence of animals as symbols throughout human history and culture is an indication that, far from being binary opposites, technology and ecology are deeply imbricated.

More than a decade before Turing’s 1950 paper, an even more tantalising glimpse of an ecology of technology appeared in his PhD thesis, written between 1936 and 1938. Almost every computer ever made is based on Turing’s automatic machine (or a-machine), where the output is determined by fixed, predetermined configuration. One problem an a-machine cannot solve, however, is whether a program or process will stop at a certain point or continue indefinitely. (The so-called ‘terminating problem’ or ‘halting problem’ also appears in ‘Old Mode’, in which ‘The only interesting bits are when an element is / selected and the next sentence is delivered on top – / This means the whole sequence might never end.’)⁵³ Over just a single page of his thesis, Turing proposes the ‘oracle machine’ (or o-machine) as a solution to the halting problem. An o-machine would permit a typical Turing a-machine to consult an ‘oracle’ during its computations—circumventing the undecidability by referring the decision to an external source. The oracle was not defined or described by Turing, except that ‘it cannot be a machine.’

Oracles have been a source of inspiration throughout the history of procedural poetry, but never more than in the work of John Cage. In 1951, Cage entered an anechoic chamber at Harvard University. Totally isolated from the world, he was nonetheless confronted with two sounds: one high and one low, which the technician informed him were the whine of his nervous system and flow of blood in his veins. Famously, the experience ‘gave my life direction,’ as Cage states in his 1989 Charles Norton lectures, which he defined as, ‘the exploration of nonintention’.⁵⁴ The closed room of the anechoic chamber inspired him to compose his mesostic poems (acrostic-like arrangements with the key vertical word running down middle as a spine or string and horizontal ‘wing words’ on either side) by consulting the I-Ching, initially via the traditional method of tossing three coins six times to generate number between 1–64, and later simulated ‘the coin oracle of the I Ching’ using a computer program called IC.

Hayles notes that Cage’s work suggests that, ‘by giving up an anthropomorphic viewpoint based on control, […] we gain a more capacious view of connection that engages us in the world rather than isolates us from it’.⁵⁵ Similarly, Conflicted Copy offers vision of intelligence as multiple and cognition as a matter of collaboration; it suggests that understanding this involves breaking out of another ‘closed room’—the conventional understanding of intelligence as located ‘in the head’. The problem with our approach to AI, Bridle says, ‘is that we have been trying to entrap a brain within the machine, when the real brain—the oracle—is outside’.⁵⁶ Many of the poems imagine forms of divination. In ‘After PDF’, the speaker notes that, ‘When my laptop froze a / moment after the video ended, and all I was able to hear / was random buzzing on speaker, I decided to listen.’ In ‘Dead Colours’, the speaker warns ‘there’s / a chance that you’ll miss the correct message / in the sky, / the one you’ve been watching / for’.⁵⁷ But primarily it is the oracular dimension of Riviere’s process that most resembles Turing’s fabled o-machine. In Riviere’s collaboration with GPT-2, each party acts as oracle for the other: Riviere submits his invention to the ordinances of the LLM, while the LLM’s dependence on human prompting means that by necessity it must ‘pause’ in its operations to ‘consult’ an unknowable other. In the poems themselves, this is represented by the space between each word, a record of the consultation between radically different intelligences: moments consisting of prompting and responding, hesitation and selection. Between each printed word there is vector space as well as the white space of the page.

Conclusion

Sam Riviere’s Conflicted Copy extends the remit of procedural poetry in the coming age of machine intelligence. By creating a space of collaboration that draws on the profoundly different ‘sensitivities to the language pool’ in poet and LLM, Riviere’s method opens up the possibility of a poetics of vector space. In this space, certainties about distinction between embodied poet and disembodied algorithm erode; while the former emerges as just as much a networked, distributed presence as any machine, the material reality of the latter—registered even in a quantifiable carbon cost for each poem—is made evident. But it is in the way the collection and its mode of composition reimagine the space between words that Conflicted Copy is most innovative, challenging the way we think about intelligence: a kind of oracular junction, punctuating each poem dozens of times, in which very different kinds of intelligence employ their distinct understanding of and approach to deploying language. In ‘Darken Fame’, we find what were once thought of as hard boundaries resolve:

                                            the world itself was

cut in two, separating your view as if by a glass partition.

At the centre however there was no apparent boundary:

an unbroken veil had blown across the air, as thick as rain.⁵⁸

The poetics of vector space model intelligence as open, collaborative, and enacted in relationship: breaking out of the closed room and into oracular space.

Competing Interests

The author has no competing interests to declare.