The crucial terms in this quotation from Maurice Blanchot’s essay ‘Reading’ are representations of thought concerning, respectively, literary artifacts in the case of ‘the book,’ plastic and/or perceptible artifacts in the case of ‘a statue,’ and human beings with the faculty of language in the case of ‘the reader’. Thoughts about these terms are being created by Blanchot’s, by my own, and now by your reading. These few words enact certain principles of representation that are pertinent to what I will say later on.

As such, these words are meant to remind us of an aspect of literary experience that renders it singular as contrasted with that of the other arts.1 Blanchot points out that the artifacts of visual art, and even time-based music and performance, as he suggests, ‘assert themselves’ (s’affirmer). Physical books, as objects, can of course be said to ‘assert themselves’ in this sense and we know, moreover, that their texts are waiting there, inside the books. But the text only actually ‘asserts itself’ if we open the book. Then, to encounter the literary artifact, we have to ‘create’ it for ourselves by reading, at every encounter. This, despite the fact that the delivery media of most literary artifacts are produced and published as finished things. Nonetheless, we only ever encounter existing literature in the knowledge that its artifacts, its readings, have been made (fait) again and again so as to assert themselves to and for many other readers like ourselves, often long before we have had the same phenomenological opportunity.2

Literary artifacts must be created by their readers before they can be related to in any manner, and before, in the context of this essay, they are able to be appreciated as the constituents of an assemblage.

Setting this rather profound insight to one side for the moment, we must bear in mind that when dealing with any kind of literary text that is purported to be a cognitive assemblage, we are not dealing with the whatever-it-is that we make when we read, we are dealing with text that has ‘asserted itself’. We are dealing with text that has a perceptible and material form and that can, as it were, interface with other non-human cognitive processes. In the case of electronic literature, the text must have asserted itself in the course of digitization, as it was turned into a digital object, effectively reasserting itself in a correspondent encoded form. Thus, when discussing electronic literature, we are constrained to accept the text as an asserted entity and defer the time of human reading, the moment when the literary artifact (Blanchot’s reading) presents itself. This constraint on our potential experience of literary text will be taken for granted throughout most of the following discussion, which is chiefly concerned with N. Katherine Hayles’s 2018 essay, ‘Literary texts as cognitive assemblages: the case of electronic literature’ and also with her underlying theoretical framework, now a wide-ranging ‘Integrated Cognitive Framework’ (ICF).3

Hayles’s article considers the possibility that all literary texts are, in her terms, cognitive assemblages, with artifacts of electronic literature taken as specific case studies. Although this does not seem to be stated outright, there is the implication in Hayles’s argument that aspects of the cognitive behaviors that she discovers and analyses within the computation underlying ‘electronic’ literature suggest that artifacts and practices of language, including existing print literature, assemble themselves with our constitutive technics in a manner that is itself cognitive, in a manner contiguous with but underlying more traditional notions concerning cognition, consciousness and, particularly cognition as consciousness. Literature is cognitive assemblage, she says. And it is notable that this more or less implicit argument from an extended analysis of such purported assemblages rhymes with certain kinds of argument from both complexity and efficacity which suggest to some researchers and thinkers that the large complex, ‘neural’ models that achieve ‘human level’ linguistic, indeed literary, behavior may offer insights concerning actual practices of language in human animals, and subsequently concerning key qualities, like ‘intelligence’ or ‘consciousness,’ that we ascribe to these actual linguistic behaviors.

Although I sympathize with some aspects of the impetus for Hayles’s project –particularly its subversion of human cognitive exceptionalism — I will argue that her framework’s general integration of a specific definition of cognition tends to disregard disjunctures — which my title reconfigures as codependencies — at certain points of interface between the proposed constituents of Hayles’s cognitive assemblages. Those that particularly concern us here are between a reading and writing human subject on the one hand and a transacting, or prompted and responding, machine on the other — actualized in the traces of language that we call ‘literary text’ — and between language as such and language as ‘represented’ in the logic of the digital.

To practice or appreciate literary art, we must attempt to take account of what language is, what language makes of us, and what we do when we understand language as such. Much of what follows will revolve around questions concerning representation. In one of her recent essays, Hayles refers to a ‘crisis of representation created by the subversion of the human aura’ (Hayles, 2023b: 274). The crisis here is one in which the exceptionalist specificity, or ‘aura,’ of literary representations that are created by humans is challenged either because they may be shown to be co-created, in her terms, by human-technological cognitive assemblages or because they are indistinguishable from asserted, usually textual, representations produced by machines in response to prompts. As subject to computation, these latter representations are, by definition, digital in their ontology. They are formal representations constructed within what Aden Evens calls the ‘logic of the digital’ (Evens, 2024).4 They are reducible to sequential patterns of computation’s one definitive difference. Any sense of ‘crisis’ arising from the identification of assemblage-generated sequences with human-generated sequences begs the question of whether or not the human-generated representations are also exhaustively described and explained by formalization.

Let’s take a moment to review some aspects of Hayles’s framework. For Hayles, computational media are nonconsciously cognitive, and can form assemblages with human and other agencies to distribute cognition and realize meaningful action.5 She defines cognition as, ‘a process that interprets information within contexts that connect it with meaning’ (Hayles, 2017: 22).6 Hayles finds cognition in everything that is biologically animate or that is created and animated by, typically, human animals. Humans themselves, for Hayles, are governed more by their biological cognitive nonconscious than by their conscious cognition, and it is certainly possible to agree with Hayles that the human cognitive nonconscious is continuous with and no ‘better’ than its instantiations throughout the biological domain. But Hayles also distinguishes and acknowledges a technological cognitive nonconscious which may be either mechanical or digital, the latter necessarily interfaced with the mechanical. She proposes that we should explore if not embrace the sense of our human post-war history as moving toward one in which the human — never for Hayles distinct from non-human cognitive nonconsciousness — becomes actually integrated into ‘cognitive assemblages’ together with cognitive nonconscious machines of its own manufacture. In the cognitive assemblage, there is contingency and creativity because the human is ‘always already’ integrated with cognitive nonconscious technics, and here Hayles can appeal to the kind of Bernard Stieglerian evolved technics that ‘the human’ derives from, for example, bipedalism and opposed thumbs.7

By contrast, I contend, with Aden Evens, that actual integration of the digital with human cognitive activity, conscious or nonconscious, is impossible due to the opposed ontologies of their respective representations. The digital can only connect with our world through our adoption of or adaptation to an abstracted representational logic that we ourselves design and encode.8 At the level of interface and interpretative possibility, Evens calls these representations ‘iconic’ in a usage that aligns both with Peircean semiotics and with the actual icons familiar to us from user interfaces (Evens, 2024: 132–134). Icons represent, through resemblance, interface elements that might exist in the actual world of human behavior and action while, in their digital world, they visualize — and they can only do this formally —formal representations that function as positivistic within an entirely rational program (however complex) that is instrumental with respect to its responses and outcomes.9

My adoption of ‘iconic’ in this context, following Evens, requires a little exposition that may help to distinguish digital representations of language from literary representations (referring back to the opening discussion of Blanchot) or those of language-as-such (whatever humans use when they read or write or listen or speak). The word is useful due to its adoption in actual computation, and its usage is properly Peircean because, for example, a ‘document icon’ is made to look like a document but is neither a physical index to the document nor a purely conventional symbol. By the same token (pun intended), for example, digital representations of words resemble words of text because they can be rendered to look like things that literate humans inscribe or print or read. But they have no physical connection to language. Materially, they are 1s and 0s in silicon. They only manifest their iconicity when they are rendered as graphic images by an output device and they require this iconicity to be humanly read (and only then do they become subject to any symbolic conventions of actual language). We must either adopt or adapt to this iconicity. The point is that, within the digital, all significant operations on these digital representations whether to analyze them into their ‘lower’ constituents (ultimately sequences of bits) or ‘higher’ into readable structured sequences (of ‘words’ for the example we are tracing) are entirely subject to the logic of the digital and nothing else. This contrasts with actual literary representations or those of language-as-such. While the nature of these representations remains a mystery, it is clear, for example, that actual words connect physically with their human users and also that these users may have conventions for their understanding that are outside any current structural linguistic system in which they also, nonetheless, operate.10

Hayles’s accounts of cognitive assemblages describe intercommunication between their chiefly cognitive nonconscious constituents as if these constituents were all more or less commensurate. She usually characterizes this intercommunication in terms of feedback loops, crediting cognitive interaction — regardless of its biological or technological, its conscious or nonconscious instantiations — as what keeps an assemblage together and gives it interpretable meaning. She goes so far as to give a special priority to ‘computational media’ as the quintessentially ‘cognitive technology [original emphasis]’ and thus ‘able to impart to a huge variety of other technologies the advantages that cognition bestows — flexibility, adaptability, and evolvability [emphasis added]’ (Hayles, 2021: 10). Any historical or situational contingency implied by these three terms cannot, however, be represented, let alone ‘bestowed,’ by the digital for which contingency does not exist. Hayles’s apparently lively, vital, animal, human adjectives are being applied to what, following Evens, can only be an instrumental efficacity, enhanced and intensified by complexities in the technological cognitive nonconscious operations of computation. Such nuance is never contingent until it interfaces codependently with our actual world. Our association with the digital cannot be one of assemblage such as Hayles proposes, it can only ever be an instrumental relation, with the actually animated party either taking a dominant role or willfully enabling and submitting to this kind of codependency. We must adopt or adapt to the icons. We cannot read them as we might wish or imagine, not in any way that would make them different, whereas we are always changing, regardless of whether or not this is due to our desire and imagination. One reason for this is because we exist in and of time. We will return to this point before concluding and this will allow us to directly, if briefly, address certain questions of narrative in literary texts with computational dependencies.

In her ‘assemblages’ essay, Hayles discusses examples of digital literary texts in terms of her framework, and we will comment on these shortly. First, we’ll present a very simple example based on a piece that I made in response to a 2025 call from Nick Montfort.11 This example has the advantage that it can be quoted in its entirety and thus read as an artifact more or less exhaustively within the confines of our argument. The following is the complete code, in Node JS, for John Cayley’s The code is saying[.js]:

This code loads itself as data, then does a little bit of simple processing, and then, for each execution, outputs two hundred and fifty-six pseudo-randomly selected words from an array that results from the preprocessing. Here is an unedited sample:

Yours truly, a human, has created, in the text-of-the-code, formal representations that do not relate to anyone’s actual linguistic behavior other than my own. They do relate explicitly, however — chiefly in terms of shared vocabulary — to thinking that yours truly has been undertaking for some time now. To put this simply, I’m reading Henri Bergson and attempting to integrate my own thinking with my reading of his. The code’s output — the generated text, although this is surely not the only text-to-be-read? — does, I think, a creditable job of providing me with some new prompts-for-thinking about Bergson’s and my own thought, but it would be difficult to claim or justify any intentionally significant linkage between such thinking and either my attempted representations in the code or the code’s highly arbitrary formal representations in terms of and in regular derivation from the code itself.12 My thinking has not been assembled with either code-text or output-text in any way that would allow their cognitive nonconscious constituents to give feedback to one another. Neither code-text nor output-text is reading or speaking to one another. And neither, as feedback, are they reading, or speaking to me. I and my thinking, on the other hand, have been reading, and speaking with both. And creating them by doing so. Recall that, as Blanchot says, I must create a reading from the output text before it asserts itself to me as, even potentially, a literary artifact.

The title of this essay suggests ‘codependent’ as a substitute for the ‘assemblage’ of Hayles’s ‘Literary texts …’ essay and her subsequent framework, the ICF or Integrated Cognitive Framework. The suggestion is deliberately somewhat playful. Nonetheless, we can refine our conception of codependency in order to propose it as a serious conceptual contender, one that better represents the relationship between human and computational text generation or manipulation when these are integrated into the production and reception of literary art.

In general, and popular psychological usage, ‘codependency’ has negative, neurotic, even pathological connotations. I do not want to dismiss these entirely because there is a quite prevalent prejudice amongst literary practitioners and critics that to be dependent on computation for the practice of literary art is not a good thing. For some of them, it is indeed a dependency. Then also, there is Stiegler’s useful distinction between adopting and adapting to technologies. Adopting resists dependency while adapting tends to require it. Thus, I would suggest that embracing the potential of codependency’s negative connotations is at least appropriately cautionary as well as more interpretively productive than the more neutral or positive ‘assemblage’.

With respect to the literary artifact as such, its assertion is in all cases dependent on an author or, if we accept arguments stemming from Roland Barthes and Michel Foucault, is dependent, in a more nuanced analysis, on an author function which is far from being a simple unitary, human creator-subject. In the special case of electronic literary artifacts, there are other obvious dependencies, and in considering these we should take on the programmer and developer’s usage of ‘dependency’. In coding, a dependency indicates that on top of the inscriptional procedures coded by the programmer-author, their resultant code may, itself, have dependencies on other software, designed by other, usually absent, authors and contributors. The electronic literary artifact has, minimally, a dependency of this kind on the software composed by its author function and, yes, often assembled from other dependencies. Here I have used ‘assembled’ since the software dependencies in question are made of the same ‘stuff’ — a homoousia of formal digital representations and logics. Designated as ‘dependencies’ these software components do integrate — cognitively, in Hayles’s terms — by sharing necessarily identical representations and, indeed, feeding information about states and possibilities back and forth on the basis of these shared identities and logics. Taking the electronic literary artifact as a whole, however, these ‘dependencies’ of software assemblage contrast sharply with the codependencies that we are trying to address and spell out. In making an electronic literary artifact I render myself codependent with a hardware-software assemblage. I imagine a meaning and I depend on this assemblage to help me create it. I adopt the assemblage, but I adapt both the assemblage and my imagination to one another codependently. My readers are enjoined to adopt and adapt to a corresponding codependency for their processes of reception and, ultimately, their reading.

My thought and action with respect to the making of an electronic literary artifact, I am saying, is not assembled with the software of the artifact, with its technological cognitive nonconsciousness, certainly not in anything like the same way that its constituent software dependencies are assembled. If there is a feedback loop it is hardly regular in any way. It is better characterized as contingent, and we’ll come back to this. To make this more concrete (and also to bring in a little Critical Code Studies close reading), in the example above my thinking/composing allows me to arbitrarily name a variable ‘whatIAmGenerating’ in the knowledge (this human’s knowledge) that it will be parsed into words for the generator’s vocabulary. The variable is declared as a String and as such serves its uses in a purely digital logical manner, nonconsciously collecting text and allowing cognitive nonconscious JavaScript methods to change its contents. But there is no conscious or nonconscious relationship between the String and its manipulation that is meaningful other than with respect to the meanings I have imagined for it, before and after it has finally been reassigned (and re-Typed) as an Array from which there is only a very small possibility that the code’s pseudo-random selection would output ‘What I am generating,’ for example, in this or any other meaningful and grammatically acceptable sequence. As author, I had to codepend on the code which then rearranged itself as data — deploying, yes, cognitive nonconscious operations — in the digital realm, entirely unconnected with my thinking. Then, for meaning to be derived from its output, readers must make themselves codependent on the code — both obviously, since otherwise there will be nothing to read, but also imaginatively, since this is the only possible explanation for any meaningful experience of reading with respect to the generator’s output or to the generator itself (as a readable text), especially in this particular case.

I am in agreement with Aden Evens when he says that the digital is entirely divorced from our actual world of contingency. It is ‘discrete, mechanistic, and inert’ (Evens, 2024: 26). Once activated, it is ‘endlessly spinning in place’ (Evens, 2024: 59). I agree that we connect to the digital only through ‘icons’ that have a mimetic handle on our actual world and that are engineered to formally define a corresponding handle within the digital. This is one reason why I prefer ‘codependency’ to ‘cognitive assemblage’. But having explored this codependency, I am now tempted to briefly propose a more generative and potentially more ‘adoptable’ approach to the human-machine interface with respect to the making and reading of electronic literary artifacts. It strikes me that something I will call ‘co-contingency’ could also be operating, and that this might help explain why practitioners do not always discount their technological investments as the ‘dependency’ that is often attributed to it by literary critics and scholars.

As an author approaches and builds a compositional system for their literary artifact, they are clearly doing so in circumstances that are contingent; that are personally, historically, intellectually, and affectively specific. The representations that they develop in these circumstances reflect this contingency, which is then, in part, formally represented in the digital as data and bespoke operations. Insofar as the code manipulates the data in terms of both built-in and bespoke operations, the author’s literary conception is subject to transformations within the digital that do not add any new contingency, apart from and optionally pseudo-randomness, but the representation of the initial contingency is nonetheless played out, with the author waiting, codependently, on the system’s response.13 At the time of response, what I am calling co-contingency is again made manifest in that, for example, the system may not have generated a response that corresponds with whatever the author anticipated or imagined. Their reading of this response — and please recall, once more, that all readings are created by their readers — is thus subject to actual contingency. Most often, in this scenario, the role of the digital — its share of ‘co-’ in the co-contingency — is overdetermined by any pseudo-randomness introduced during the transformative operations, as it is in the simple example that we’ve been looking at so far. But is it worth noting that deterministic digital operations can also render authors and readers subject to co-contingency when the author is not sure of the effects that will be produced once their composition of bespoke and built-in operations is applied to their particular curation of data. Many pieces of literary art with computation by Nick Montfort have this characteristic and perhaps the most spectacular example of this kind of work is his collaboration with Stephanie Strickland, Sea and Spar Between, which is discussed in Hayles’s essay and also here below.

The other works discussed in Hayles’s essay are Nick Montfort’s Taroko Gorge (and its ‘remixes’ by other authors), M. D. Coverly’s Pacific Surfliner, Johannes Heldén and Håkan Jonson’s Evolution and John Cayley’s riverIsland — the latter only obliquely since Evolution substituted for it when the essay was published, after its paper presentation.14

Before discussing her examples, Hayles sets out a number of questions that she plans to address, ‘How is creativity distributed between author and computer? Where does the nexus of control lie, and who (or what) is in control at different points? What kinds of selections/choices does the computational system make, and what selections/choices do the human authors encode? What role does randomness play in the composition? Is the main interest of the artistic project manifested at the screen, or does it lie with the code?’.

Responses to these questions have already been suggested in our discussion so far, with its critical reading of Hayles’s framework and its representation of various contrasting positions. Hayles’s first and perhaps most provocative summary question evokes ‘creativity’ so as to allow it, relatively unproblematically, to be ‘distributed’ to a machine. In her discussion of her first example, the slot algorithm of Taroko Gorge, she is quick to suggest that this very simple algorithm is one way to ‘enlist’ the ‘computer as co-author’ (Montfort, 2009). This example is almost as simple as The code is saying[.js]. ‘Slots’ are positions in Montfort’s poetic structure which may be filled by randomly selected items within per-slot arrays of orthographic words curated so as to be able to fit their slots. ‘Co-authorship’ in this case amounts to a small number of technological cognitive nonconscious pseudo-random choices. It is difficult to recognize this as a ‘distribution of creativity’ unless one has already accepted Hayles’s implicit suggestion — and this is the point I am making — that the human author is always already integrated into a cognitive assemblage that constitutes their (using this pronoun for the assemblage) and, by implication, our understanding of ‘creativity’.

I’ve already set out an alternative critical position. Any proposed cognitive ‘assemblage’ of human reading and representational strategies with corresponding digital representations cannot be ‘integral’. They cannot amount to the instantiation of an ‘integrated cognitive framework’ because digital representations can only be defined on the basis of abstract, formal — Evens’ ‘iconic’ — definitions. The human author’s reading and representational strategies can only ‘feed forward’ certain positively defined and instrumentally effective characteristics to the digital. We are, therefore, entitled to reserve to human readers any play of the representations’ ‘meaning’ and ‘interpretation’ as the manifestation of ‘creativity’ in both a more general and a more philosophically coherent understanding and usage than Hayles’s.

Before going on to address Montfort and Strickland’s Sea and Spar Between, Hayles briefly compares the Taroko Gorge program with the monoglot occupant of John Searle’s infamous Chinese room (Searle, 1980). Searle maintained the occupant’s rule-based lookup did not amount to knowing a language, such that even an ‘assemblage’ with a human in it could not be said to ‘know’ or be able to converse in a language ‘intelligently’ or ‘meaningfully’. For Hayles (somewhat controversially I think), the cognitive nonconscious processes of Taroko Gorge are ‘far more detailed and complex than the ‘rule book’ that Searle imagines’. The program, she says, ‘knows’ data structures, and syntax parameters for word selection; it ‘knows’ display parameters and animations; it ‘knows’ its randomizer. But this purported ‘knowledge’ is precisely what the author of the code, Nick Montfort, has more or less consciously designed and composed or, in the case of the pseudo-randomizer, has found built into the programming language by other designers. Hayles says herself that ‘the program understands nothing about the semantic content of the vocabulary from which it is selecting words’. More to the point, there is no evidence of any kind of operative cognitive nonconscious processing that might produce a feedback loop between the program’s operations and either its own design or our reading of its output. And any input to the program, in this case, amounts to other human authors composing new arrays for word selection within the constraints of the existing design, based on their imaginative speculations concerning semantic domains that the intervening authors believe might yield novel semantically implicated literary effects. That is, rather than existing in assemblage, human co-authors look to possible output responses from the program codependently based on the adaptation of their imaginations to the program’s design.

In Sea and Spar Between and its companion operative code comment essay, ‘cut to fit the toolspun course,’ we experience the effect of computationally generated literary scale and undoubtedly also the strong impression of a technological literary or pseudo-literary sublime, as Hayles acknowledges (Montfort and Strickland, 2010; 2013). But the arrangement of text that the program generates — at all levels of structure — is deterministic, and its poetics is entirely designed by its human co-authors. The scale in terms of the number of quatrains and their distribution is exhaustively a function of the mathematical permutation of differently categorized sets of elements: vocabularies and syntactic patterns derived, by Montfort and Strickland, from the source texts of Emily Dickinson and Herman Melville. The result is a tour de force of deterministically generated electronic literature and also of literate programming, in that the commented code — composed, again, by the co-authors — can be humanly read as an essay on the project interleaved with operational JavaScript, CSS, and HTML rendered more generally readable by the essay comments (Knuth, 1992). But given that there is no pseudo-randomness in the manipulation of text, in this example there is even less involvement of technological cognitive nonconscious processes for the generation of the display-to-be-read than there is for Taroko Gorge.

Hayles’s discussion of M.D. Coverly’s work is couched in terms of ‘influence,’ with Harold Bloom’s theory as backdrop and ambition-for-canonization — as writerly motivation — that Bloom proposed compared with the ‘push-and-pull of software as literary influence,’ that Coverly attests. ‘Software’ here is chiefly digitally afforded access to audiovisual media as literary inspiration and as integrated with both Coverly’s compositional strategies and the evolving potential of non-conventional publication, ‘only possible because of networked and programmable machines’ (original emphasis). Hayles says that such facilitation of new creative practices, which would otherwise not be possible, ‘is the large sense in which computers are our symbionts’. In this context and with respect to this example we do not, however, seem to be addressing any kind of cognitive assemblage established by conscious or nonconscious processes operating or feedback looping between human author function and the media-wrangling software that the author, in this case, simply uses. This example seems tangential to both Hayles’s argument for assemblage and my critique.

As I mentioned, Hayles’s essay was originally a paper presentation and used my own riverIsland as an example text, later substituting Johannes Heldén and Håkan Jonson’s Evolution (Heldén and Jonson, 2014; Cayley, 2008; 2003). This was, in part, because Hayles had written about riverIsland previously and because its QuickTime formats were by then no longer Apple-supported such that the work was (and is) difficult to consult for Hayles’s readers in particular.15 The substitution was expedient for other reasons since Evolution’s sophisticated application of Markov models provides us with a work to explore that has a direct relationship with other work engaging contemporary large language models (LLMs). Before commenting on Hayles’s reading of Evolution, however, we should note briefly — based on the PowerPoint of Hayles’s original presentation — that she read riverIsland in a very similar manner to the way in which she read her previous examples, ascribing ‘knowledge’ to the computer with respect to the way the program integrated movie control and text display, to the operative ‘textual distance’ between texts comprising the work, to the integration of audio and visual components, and to a movie that morphed through logographic and letteral forms.16 Although the circumstances here are far more complex, as author I can attest that all these elements were author/coder designed and built into the architecture of the work. Once set in place, the cognitive nonconscious choices made at time of execution are both regular (with some pseudo-randomness) and anticipated, and do not feed back or forward in a manner that is ‘assembled’ with the ‘knowledge’ that I designed into the system or any knowledge of the work’s readers.17

In Evolution we have a generative electronic literary text that is made from a Markov model combined with a good deal of peripheral software mediating its output.18 Training data for the model is a curated corpus of Heldén’s writing. The relationship between input and output is exactly the same as it would be in the case of current LLMs, the main difference being that Markov models contain, at most, a few dimensions in their ‘vector spaces’ pertaining exclusively to the relative frequency in their corpora of (usually) word-sized tokens and the relative frequency of token sequences of various lengths.19 Evolution differs from more conventional models in that it also treats spans of ‘white space’ as tokens. These particular tokens are derived as data from the same corpus; they are simply the white spaces of Heldén’s writing. This property of the work highlights the way in which aspects of linguistic artifacts have been selectively represented by computation. White space is constitutive of poetic language but both linguistics and computer science is unclear about its status as language. The ‘word processing’ of computer science regularly collapses any white space to a single digitally abstracted space character.20 Evolution’s model does not, and this is a brilliant, all but unacknowledged, innovation.

Nonetheless, these usually disregarded tokens do not change the fundamental structure or operation of the Markov model. It is ‘trained’ in a relatively simple process that builds a digital representation, a probability table of token sequence frequencies — here including white space tokens — and, once built, this data structure is Evolution’s particular, and essentially static, model. A ‘prompt,’ to a Markov model like this one, is simply a request to ‘start’ with a particular token or sequence of tokens. The model responds with the token that would most frequently follow the last token and/or the last token sequence that has been output and/or prompted. Adapting Evens on LLMs, the chief component of Evolution is a database in the form of a Markov model.21 You may prompt it to provide you with word sequences related to your prompt that are marked as relatively frequent in the model’s structured database of Heldén’s writing. As implemented, Evolution is activated by starting with a randomly generated query after which peripheral software continually feeds the preceding tokens back into the system as new prompts, generating textual sequences until given a command to stop.

Everything else is interface or what developers call ‘chrome’. Any proposed ‘co-authorship’ by technological cognitive nonconscious components must be attributed to the model. Again, my reading is that there is no ‘assemblage’ that would be brought together by shared cognitive feedback loops between human author and the software-prompted model, nor any such cognitive transaction between reader and model as output/responses are generated. Instead, I say, both author and reader functions render themselves codependent on Evolution — as an animated literary artifact — in order to experience their readings. Heldén flirts with the idea that Evolution takes over from him as the generator of his writing, and I say that this is simply the ultimate expression of a codependency that he has set in place.

In the authors’ exposition of their work and in Hayles’s commentary, much is made of the part played by randomness. Hayles and Heldén reference John Cage and an approach to authorship or art making that adopts chance (I am consciously using Stiegler’s terminology) in a spirit of submission to the world’s contingencies and as abdication of mastery and self-regard. I would not want to deprecate this mode of art making in any way. But I do not see randomness as an operation of technological cognitive nonconscious processing, particularly within the digital. In this realm, it would be difficult to claim that randomness ‘interprets information within contexts that connect it with meaning’ (Hayles, 2017: 22). Rather, I see it as an aspect of the co-contingency — as I have introduced it earlier — that emerges from the artist and reader’s codependency on systems, like Evolution, where randomness operates. Indeed, this helps to actualize the type of relationship to chance in art making that Cage, Heldén, and many others, including yours truly, value and undertake.

The randomness in Evolution is both nuanced and elaborated in that it is used to make a choice of operation with respect to the output of next tokens, and because ‘seeds’ resetting computational pseudo-randomness are generated from otherwise unrelated datasets that appear to have been selected for their semantic and affective resonance. Tokens may be added, subtracted, replaced, or subject to ‘white space substitution’ (Heldén and Jonson, 2014: appendix 10). So-called ‘natural’ seeds for pseudo-randomness are generated from datasets associated with ‘atmospheric data, visual imagery, space observations and popular culture’ (ibid). There also seems to be a mechanism by means of which generated text is fed back into the model so as to reconfigure — metaphorically evolve — its vectors of collocational frequencies. None of these features challenge my reading of Evolution’s operation as that of a digital artifact having only a codependent relationship with the author’s or our own experiences of literary reading as and when we engage with it. Letting the output change the model feeds nothing back or forward to the user, and the processes of further generation by the model remain unaltered. The dataset-governed selection of seeds for computational pseudo-randomness seems unlikely to have a legible effect on Evolution’s behavior and, if anything, this is simply the further mark of a Cageian submission to chance operation on the one hand and, on the other, the solicitation of meaningful coincidence by the authors, who would like us to be thinking about the atmosphere, visual imagery, space, and popular culture as we read and as Heldén did, presumably, when he composed his writing before Evolution.

Despite conceding, at the outset, that we’d be dealing with asserted text rather than critically co-imagined actual readings as they are created by human authors and readers when they encounter such texts — despite this, we have glossed over the obvious fact that the text in and of electronic literature, including large language model output, looks, on the textual face of it, exactly like the text in literary books or like the text that we word process into a document such as this one. We take the text in books to be the recorded traces of an author’s language: what they said. We read these words in a context that is not just textual, imagining a meaning for the text that we trust will resonate with the author’s intended meaning. This amounts to saying that the literary text — and any text of language as such — is a representation not only of structural difference but of a relation between reader and author functions and between both of these functions and the world. The text of the digital cognitive nonconscious, on the other hand, is comprised of formal representations subject to the logic of the digital. Within the digital, any significance these representations have is entirely determined by immediate context in a manner that can be exhausted by a purely formal, structural analysis. And within the digital, these representations have no affect at all since digital artifacts are not embodied in an actual world which would allow them to develop the kind of affective structures that emerge from our embodiment in a world that both nurtures and threatens our embodied well-beings.

In her later writing on LLMs Hayles touches on such distinctions by suggesting that we take into account our human being in a particular umwelt, and the technological cognitive nonconscious of LLM computation existing and acting in a more or less related umwelt of its own.22 This seems to me simply to be an attempt to collapse the real distinction that I have just set out. Can an inert, entirely rational and instrumental process, with no determinate body, ever be considered to exist autonomously in a world and to have a construable umwelt of its own?

Regardless, the output of electronic literature, from Taroko Gorge through Evolution to LLMs takes the form of text. A seductive but false (hence its conclusion is marked with a *) syllogistic argument is provoked. The output of electronic literature takes the form of text. Human language takes the form of text. *Therefore, the output of electronic literature is human language. For linguists and computer scientists there is no problem. Arguably, the object of linguistic science is edited transcription. In other words, it is text subject to (grammatical) ‘acceptability’ in forms, usually sentences, that can be distinguished from those that are unacceptable and that are marked as such (with an asterisk *) in the written archive of linguistics. Computer science and computation in general necessarily takes on this view of language since it can only ever deal with transcriptions that can be subjected to formal, digital representation. And the research-agenda development of LLMs is inevitably subject to assessments based on the criteria of linguistic science. How, they ask themselves, can machine learning researchers make LLM output ever more ‘acceptable’ in such terms?

So, for many of us — and for some of the most culturally powerful amongst us — language simply is its traces. What else do we have of it that we can render as data? There must be a meaning to any arbitrary identification of a transcribed, usually textual, linguistic form with any other instance, in any context, however generated.23 In any case, Hayles’s technological cognitive nonconscious can only make and operate on these kinds of idealities, and thus the threshold between human cognition and the technological cognitive nonconscious offers no resistance, at this point in Hayles’s thinking, to a materialist, scientistic understanding of language.

Someone taking such a position must argue that the ‘linguistic’ transcriptions in code or output — as in The code is saying[.js], for example — are acting within ‘language’ and are ‘connected with meaning’ as such. These transcriptions have been ‘assembled’ with the code, they say, and then cognitive nonconsciously involved in its execution. Well then, any meaning in the resulting output can be ascribed as much to the machinations of the technological cognitive nonconscious as to its only partially collaborative human author.

Opposing this, we say that the representations instantiated in digitally transcribed text are not ontologically the same — regardless of their having the same form — as those produced in the language of author and reader functions. In order to derive meaning from the output of electronic literature human author and reader functions must codependently behave as if the representations are the same and as if this output is language. But nothing of the technological cognitive nonconscious operations of the electronic literary artifact is ‘connected to meaning’ so long as they remain within the digital. If these operations are indeed cognitive, they are cognitive in a manner that is very different and considerably impoverished as compared to other, basically biological, cognitive nonconscious operations in ourselves and in all the other animate umwelts of our world, including our inner cognitive nonconscious linguistic operations.

Digital linguistic transcriptions are, precisely, ‘icons’ in Aden Evens’ sense. They ‘look like’ things in our actual world of language but function only as positivist formal representations in the digital. We are codependent on these icons to make the digital do any kind of work with language. But taking this further, and to a less bleak circumstance than one of absolute disjuncture, we can acknowledge that the iconic characteristic of digital linguistic elements is also characteristic of all linguistic traces. Words, for example — to address the element of language that most easily comes to mind in our general linguistic understanding — offer us points of inflection by means of which we may fold literary text onto and over our thought and imagination, and vice versa — in a Giles Deleuzian sense of ‘the fold’ (Deleuze, 1993).24 This is an essential part of reading/understanding. When we attempt to read and interpret the outputs of electronic literature, all we can do is readopt the codependency that we initiated as author and attempt, as reader, to fold the output onto and over the thought and imagination we anticipate, hoping to bring new, perhaps unexpected, meanings to this experience of reading. Success in this undertaking is what I am proposing to call co-contingency, a way for the otherwise disconnected machinations of the digital to play a part in our actual and contingent world.

The case of LLMs as electronic literature in this critique of Hayles requires special attention at this point. From Taroko Gorge to Evolution we are dealing with what I call, following Daniel C. Howe, heuristic algorithm (personal communication). We are acting as if the digital linguistic representations are language and we understand what our algorithms are doing with these representations. We could reiterate these processes with other linguistic ‘data,’ meaningfully compare the results, and perhaps even make some discoveries: aesthetic, linguistic, whatever. For Hayles, the LLMs seem to represent nothing more than an advance in the progress of literary cognitive assemblage, an opportunity for the technological cognitive nonconscious to take us further beyond the exceptionalism of our literary representations and their aura. I say, with Aden Evens, that an LLM is ‘a database in the form of a neural network’ just as Evolution is a database in the form of a relatively sophisticated Markov model (Evens, 2024: 189–90). The difference is that I can understand and anticipate the effects of the responses from a database like Evolution whereas the layered vector spaces of LLMs are vast and highly dimensional beyond human interpretability, for now and perhaps, according to some researchers, in principle. What can be demonstrated — with respect to the questions of digital and actual representations that we have been discussing — is that LLMs ‘cut themselves off from any shared contingent representations of language to resist sayability and embrace abstraction at the very launch point for their datafication,’ as I put it in a creative piece exploring this problem (Cayley, 2025b). Whereas Evolution builds a Markov model out of things we recognize as words and, with brilliant unconscious insight, also out of the poetic spatialization of ‘white space,’ the LLMs that now intervene in nearly all our practices of language are built from algorithmically derived tokens, from character-sequence tokens that are shorter than most words and often non-lexical. The outputs that we read from model interfaces are reassembled strings composed from these tokens, not from words. In terms of the type of ‘folded’ reading that I outlined above as a way of codependently re-integrating the reading of electronic literary output with the reading that we undertake when reading books, for example, the question of how to fold is moot. We see the ‘words’ of the LLM output but are they really there? How should we fold this output of tokens onto thought and imagination which has been, since literacy at least, closely and meaningfully related to words?

In the case of the LLMs our codependency must needs run deep and, if it is to be generative and productive, come into actual relation with reading that, for us humans, is rooted, ultimately, in our bodies. It is not simply the fact that LLMs are trained in terms of abstract formal tokens. As Evens helps to show, within the digital all subsequent, digital, logically derived, representations — however obscure, whether or not they are sealed, hermetically, from human interpretability — are known to be isomorphic and of the same abstracted digital substance. Unlike analogous structural layers in the actual world, where, for example, walls may not be regularly or exclusively built and encoded in terms of and in the same substance as bricks, ‘all of the digital layers are isomorphic, each tied deterministically to the others, such that there can be no substantive tension, no disagreement in these layers, and ultimately nothing available on one layer that isn’t strictly captured in the others’ (Evens, 2024: 148, original emphasis). In the actual world of reading, however, we’re not talking about walls, we’re talking about whatever it is that we create when we read, the literary artifact that we make when reading and when writing. It is impossible for yours truly to believe that this whatever-it-is, the literary artifact, is of one substance and has been digital logically derived from grammars and from ‘words in my mind’ even if I was convinced that ‘words’ were somehow ‘in there’ at all. Think about it. In my own processes of thought, the representations and my sense of them have nothing like the extent or shape of any kind of transcription and the way in which they operate within me is significantly nonconscious and only very rarely comes across as ‘computational’. And what does arise for me usually does not cohere rationally in terms of posits. I prefer to imagine words, indeed, to imagine language moving within the time and space of a body-mind that Roland Barthes describes as ‘extended outside laws which proscribe contradiction’ where reading ‘is the permanent haemorrhage by which structure … collapses, opens, is lost … leaving intact what we must call the movement of the subject and history: reading is the site where structure is made hysterical’ (Barthes, 1986: 42–43).

It is hard to conceive of any cognitive nonconscious feedback loop connecting and ‘assembling’ what happens in the model with what happens during my reading of its output. In our electronic literary context, I enter into a relationship of codependency with the model when I give it my prompt. I’m not assembled with it. It takes my prompt and turns it into token sequences that are algorithmically abstracted from the actual language of my prompt. The model is not ‘feedback looping’ with it. Nothing derived from my prompt is ‘connected to meaning’ during processing; it’s just connected to a database of text and other data in an obscure and complex manner. The same kind of relationship persists when I receive a response. Once it responds, in fact, the model is codependent on me to ‘fold’ the text that it outputs onto and over my memory of literary artifacts and, if this is possible, allow me, the reader, to create a new one.

The possible creation of an actual and actually meaningful literary artifact from our encounters with electronic literature requires, in this analysis, a codependency that cannot, I argue, be interpreted as the kind of cognitive assemblage set out in Hayles’s framework. All digital components of this relation I take to be inert, rational, and instrumental, and as such impoverished with respect to the generation of new meaning and in terms of any ‘connection’ to meaning. But the relation of codependency is one that is within our gift to develop and elaborate. We can, for example, choose between adaptation to and adoption of, recalling Stiegler’s distinction. As author functions, coming to work with electronic literary systems, we also embrace the co-contingency of this relational moment, and I have suggested above that, along with codependency, this plays out and rejoins the actual when the system produces a text to be read. During all of this augmented writing and reading process as a whole, the critical default is to place the system under the sign of the attributed, named author function: for example, riverIsland by John Cayley, Evolution by Johannes Heldén and Håkan Jonson. Part of the rhetorical appeal of Hayles’s ‘cognitive assemblage’ is that it retains a sense of agential integrity while distributing — nonconscious, not necessarily attributed — cognition. In this sense, assemblage can appear as a concession to the human cognitive exceptionalism of the named authors.

I believe that there is another mode of practice in these circumstances that has, in the examples so far discussed only been operative and generative when chance as co-contingency is invoked, when Heldén or, indeed, Hayles as critic, or when I myself invoke a Cageian submission to the arbitrary contingencies of our datafied ‘preparations’ (as in ‘prepared piano’) and to actual ‘chance operations’ in the systems addressed. In all the examples above, we more or less conceal what I read as a non-assembled disjuncture between text that a human author function writes or reads and text that is computationally processed and then left to be read as if it were no different. In Kyle Booten’s recently published Gyms, the kinds of relations for and with electronic literature that have emerged from this discussion are not only acknowledged, they are explicitly developed into a poetic practice (Booten, 2025).25 Booten’s ‘gyms’ set up linguistically generative computational transactions that swap the expected roles of prompter and respondent.26 The systems still have to be ‘started up,’ of course, but Booten’s gyms software generates prompts for human reader/writers to which the latter are invited, indeed challenged, ‘athletically,’ to respond. The responses may then also be fed back into the systems, resulting in a poetic dialog. In their current published form, ‘Green text was generated by a gym. / Black text was generated by the subject,’ the latter, in this instance, being Booten (30). As such, the distinction between computationally generated text and the human subject’s response is clearly marked throughout. More to the points we have discussed in this essay, the supposed mutuality of ‘co’ in our proposed codependency is realized: not only is the human subject codependent on the computational system, the system codepends on a responding subject. With respect to co-contingency, any contingent or arbitrary conditions in the system are made actual in the software prompt, such that further actual contingencies brought to bear by the human writer — what Barthes calls ‘the movement of the subject and history’ — may be integrated, actually and contingently, with whatever the system proposed.

In this mode of electronic literary practice, it becomes less of a concern to assess or judge the computational system — particularly any supposed ‘human-like’ or ‘human-level’ qualities — because, at least initially, the system doesn’t pretend to replace, or imitate, or subsume a human language user. Codependency is acknowledged, but any instantiation of ‘assemblage,’ along with otherwise foreclosed questions concerning the ontological status of the computationally generated text and its representations, is productively side-stepped. The human subject responds as a human rather than prompting its computational respondent to behave as if it was human, to behave as other than what it is. The system feeds co-contingency forward to the actual reading and writing of ‘the subject’. Booten’s Gyms invert, and literally subvert, the order of entities which transact in terms of prompt and response. But it is worth bearing in mind that the underlying relations governing similar transactions as they are now conventionally, habitually transacted, especially with LLMs, are fundamentally the same, it’s just that our appreciation of the stakes — artistic and ethical — for certain kinds of computationally codependent engagement in literary art — a highly valued human activity — are clarified rather than eroded.

It is notable that in all of this and Hayles’s discussion of electronic literature, there is very little if any discussion of the literary artifacts’ content. In Hayles, a brief passage of close reading from Taroko Gorge consists in highlighting the ‘natural landscape’ vocabulary of Montfort’s original version and contrasting this with the ‘urban’ vocabulary of Scott Rettberg’s first variation, Tokyo Garage. But this pertains to an entirely human cognitive operation, nothing to do with the program’s technological nonconscious. Neither Hayles nor I have provided a reading of Heldén’s poetry or of any particular Evolution poem. We haven’t made or found a Deleuzian fold or taken the time to create comparative readings in Blanchot’s sense. Frankly, neither Hayles’s essay nor this critique is the place to do so. We are chiefly concerned with differing views and theories of the linguistic and formal representations and operations that are encountered in electronic literature and what this means for reader and author functions.

Something that we can say with respect to content, however, is that none of the examples discussed, apart from M. D. Coverly’s work which I signaled as tangential to Hayles’s arguments and my own, is an instance of longform or even shortform fiction. The origins of electronic literature as a field revolve around hypertext but the poetics of this form are chiefly at the level of the paragraph or short section (such as might be marked by a ‘space break’ in print). Typical hypertext nodes are author function composed, not generated with computation. In a sense, the ‘assemblage’ framework is more pertinent to hypertext since both author and reader have an articulable stake — the terms of which could be exchanged via cognitive feedback loops — in what to choose to make readable or to read next.

The electronic literary texts that Hayles and I have addressed are more poem-like than short story- or novel-like. We are debating the theory of their poetics at the level of the verse or sentence and below. Evolution as an asserted artifact is a book of code and output with documentation. As a website it is at least book-sized if not all but infinite in linguistic extension, but it tells no story. As a model, I have noted it to be a precursor of the LLMs. We’ve discussed Evolution and gone on to critique Hayles’s framework with respect to LLM text generation without, yet, taking up the question of large model generated narrative or longform fiction. There are a number of published instances of this already and the models clearly have no problem generating any quantity of prose.

To properly treat the question of narrative and LLMs is far beyond the scope of this essay but on the kind of philosophical grounds that we have so far trespassed, there are a couple of things that can be said here. I’ve agreed with Aden Evens that one of the characteristics of the digital is that it is inert. It does not move unless given instruction and energy to do so, and the movement it can make is entirely governed by the logic of the digital. Digital representations move from inert state to inert state (however quickly) and none of these states have any significant relationship to time in the admittedly problematic understanding of entropy-driven time to which biological entities are markedly subject. Just as there is no arrow of time in Einsteinian physics (and technically no need for any kind of time variable in its descriptive equations) there is no time in the digital (Rovelli, 2018). The problem of narrative for LLMs is not only the lack of a ‘world model’ but there is also the fact that the models do not have time. Trained, they are snapshots, databases in a particular state and provided with an extraordinary hyper dimensional vector space of possible combinations but … no passing time.

When I address you, even as I am now, in this writing, I do so knowing that I put these words and thoughts together at a particular moment in my life to which I cannot return, and I know that when you read this, you will do so at a particular moment of yours. This is narrative itself. It is something of which I can make literary art, but a large language model cannot.

Interestingly, if we retain an LLM transactional scenario in which the model is simply prompted to generate a narrative or narrative fiction, it is easy to see that the model, for the philosophical reasons already set out, is not the kind of entity for which this could be in any way connected to the narrative meanings of beings who live in time. Any success in such narrative art could only ever be a function of peripheral intervention at the prompt or response phases or due to guidance and corrective coercion in chat dialog with a human subject. On the other hand, if the model were treated as a ‘gym,’ it becomes quite reasonable, if not exciting, to imagine it prompting ‘the subject’ with narrative-like fragments and then to see and read how the subject was able to engage, resolve, or extend the prompted potential until it was worthy of literary history, or of history, plain and simple.