A Portrait of the Artist as an LLM
Writing after the death of the (human) author
Welcome to Ever Not Quite—Essays about technology and the human.
My last post about attention ended on the subject of reading, which happens to be where this next one picks up. There’s been a lot of recent discussion about the unique intervention of LLMs in the domain of writing and language. This, for now, is my contribution to that conversation.
As always, thank you for reading. If you’d like to support my work, consider subscribing and sharing this essay around within your circles. It’s the best way to get it in front of new readers.
“Don’t you see that with volume alone we’ll completely overwhelm them! This machine can produce a five-thousand-word story, all typed and ready for dispatch, in thirty seconds. How can the writers compete with that?” Roald Dahl, “The Great Automatic Grammatizator”
Indulge me for a moment in a quick thought experiment. Suppose that sometime in the future, LLMs get very, very good. Suppose they get so good, in fact, that their output becomes indistinguishable from even the best human authors, producing entire novels that are well-wrought by any literary standard: they’re structurally balanced from beginning to end, have interesting characters with complex motivations, and handle themes of humane concern deftly and with apparent care and sensitivity. And beyond considerations of artistry, imagine that publishers, market research in hand, excel at forecasting which of these books are likeliest to find mass appeal, and use these figures to fashion prompts for the mechanical scriptorium. In effect, two of the publishing industry’s oldest problems—the scarcity of talent and the caprices of the market—are effectively solved with the malleable magic of information.
Now imagine that none of this is a secret: everyone knows these books are algorithmically-generated, but attitudes towards the use of AI in creative fields have softened to the point that none of this much bothers most people anymore. Perhaps some murmur of nostalgia lingers in the thought of reading a book that was written by a human individual who alone conceived each detail and chose every word, and this means that classic novels are no less admired than they are today. But people have grown up around these sorts of authorless texts, and any sense of real unease or scandal has long since abated. Algorithmic authorship has become as commonplace as CGI in movies, which was itself not uncontroversial when it started to become widely used thirty years ago. People know about the before-times, back when books were the laborious product of human striving and genius, but few actually remember it. Perhaps publishers will spin up AI-generated ‘authors’, each with its own style and sensibility, cousins to today’s synthetic ‘actors’ like Tilly Norwood which film studios are just beginning to push in earnest. Publishers know that, like today’s movie audiences, the reading public wants something entertaining, interesting, insightful, and perhaps occasionally profound, and these new books are all of these things to varying degrees—and now they can be published faster and more cheaply, all without compromising literary merit. The whole thing looks like a veritable abundance agenda for the Republic of Letters.
And let’s imagine, in one last morbid surmise, that this occurs far enough into the future that all serious human writers have since died, leaving behind none who has ever authored a work of any real ambition or sophistication. Writing has altogether ceased to be a way to make a living, just as you won’t find a career working as a “computer” as you once could; such a thing, after all, would make about as much sense as a NASA engineer performing calculations with pencil and paper without the aid of a calculator: far slower, more expensive, and less accurate. The human author is as dead as the human computer, and a new profession of ‘prompt engineer’—perhaps itself some kind of AI agent—has assumed its place, goading the machine into generating 300-page novels in just a few seconds.1
What exactly has been lost here, and what gained? Books are still being produced—by some lights, great books, and at a much faster clip than ever before—and readers have more access than ever to anything they could possibly want to read. It hardly matters which of two worries seems the more urgent: that AI really will produce good writing, or that the writing it produces will be taken to be good by an undiscerning population which has ceased to cultivate a critical taste for the written word. Just as surely as you are reading this, I have a personal stake in these questions, as does anyone for whom writing is something more than a hobby or a mere tool to be blithely turned over to an algorithm offering superior functionality for less cost and effort. You might dismiss this scenario as unfeasible, notwithstanding the astonishing achievements of modern LLMs, but here I want to treat the rise of sophisticated language machines as an invitation to wonder about the phenomenon of authorship.2
If you’re put off by what I’ve just described—which, to be clear, is intended not as prophecy but only to highlight what’s at stake—it’s likely because it betrays a standard intuition about what language is and what writing accomplishes. Among its most vivid expressions was recently voiced by the South Korean novelist Han Kang, winner of the 2024 Nobel Prize in Literature. To read and to write means “[f]ollowing the thread of language into the depths of another heart, an encounter with another interior”, Han said in her acceptance speech in Stockholm that year. It involves “[t]aking my most vital, and most urgent questions, trusting them to that thread, and sending them out to other selves.”3
This view of textuality is common, even if it is here articulated with unusual grace: the text is a thread, a vehicle which connects one self with another, granting unique access to someone else’s inner life.4 A written document is a powerful technology of intersubjectivity, one which facilitates an encounter between individuals against the constraints imposed by intervals of time and space. It’s how Cicero, writing from Rome in the waning days of the Republic, could reach across provinces and centuries to contact Petrarch as he rummaged through the Chapter Library of Verona Cathedral after more than a millennium of cultural and political transformation. (Petrarch, in fact, would return the correspondence of his long-dead interlocutor, beginning one letter in 1345 with the words “I fear that my last letter has offended you…”, so intense was his sense of connection to the Roman writer, nearly fourteen centuries dead.)
Writing is action at a distance: from the author’s point of view, the reader is absent, yet vividly imagined; every word is chosen according to the effect it is expected to have on the text’s future audience.5 From the reader’s point of view, the author is absent, yet ‘heard’; the text reveals a mind stabilized, preserved, and splayed out on the page in a way that the ephemerality of oral communication never allows. But the scenario I’ve outlined above violates the intersubjective promise which until very recently every text could be assumed to make. Now, there is no longer any subject, no ‘who’ awaiting the reader on the other end of the thread, only the output of a weighted dataset of a trillion embedded vectors—information, aggregate and bodiless. The thread of the text now leads nowhere, and the implicit pact between text and reader has been breached. Whether or not you believe AI could ever be conscious, it’s clear that algorithmically-generated text does not reflect the inmost revolutions of a subjective interior similar in structure to your own. The reader is denied access to anyone we could call ‘author’ who might possibly be reached on the thread’s other end. All that remains is the text itself.
“For a text to convey its message, it does not matter whether the author is dead or alive.” Walter J. Ong, Orality & Literacy: The Technologizing of the Word
By the time Roland Barthes published his seminal essay “The Death of the Author” in 1967, its basic premise had already been commonplace within literary studies for several decades.6 You could even trace its anticipations as far back as the nineteenth century’s insistence on “l’art pour l’art” (“art for art’s sake”), when literature and the other arts began to assert their autonomy by disavowing didacticism and any subordination to a social purpose which might justify their existence. With the rise of New Criticism in the 1930s and ’40s, references to the accidents of biography or history which happened to contribute to the text’s production came to be seen as distractions from the meaning that is conveyed by the words themselves. Barthes merely took up this idea and pushed it a step further.
What is the source of meaning in a text? Traditionally, the answer had something to do with the person who wrote it: the text means what its author is trying to say, and our reading is successful to the degree that it reconstructs the author’s true intent. It’s not by accident that the words ‘author’ and ‘authority’ both derive from the same Latin root: ‘auctor’, a source or originator. When we follow the thread of the text in search of a ‘who’ awaiting us on the other end, we’re looking for a figure endowed with the power to authorize the terms of the text’s legitimate interpretation. If we conclude that the text means something different from what the author intended, this can only be because we have mis-interpreted. In this way, meaning is subordinated to the author: as Barthes puts it, “[t]he explanation of a work is always sought in the man or woman who produced it”.
But Barthes thinks this is a mistake. ‘Authorship’ is a concept which we impute to the written word, not one which belongs to it inherently. The ‘author’ is “a modern figure”, and therefore an artificial one, a product of contingent and unstable conceptions of subjectivity as well as practical considerations concerning legal copyright which arose with the technology of printing—all of which are diversions from the textual tapestry of interdependent meanings that are fixed right there on the page. Jacques Derrida’s famous declaration from around the same time captured the idea with decisive clarity: “il n’a pas de hors texte”—there is nothing outside the text.7 The written word thus ceases to be a window through which we can access something (or someone) else, and now becomes the surface upon which we discover the creative performance of language itself. “It is language which speaks,” Barthes announces at one point, “not the author”. The time has come, he concludes, for the specter of the author to pass away.
Every text enters the world an orphan. The written word, Plato’s Socrates had long ago complained, is dumb and impassive in the face of questioning, helpless without the living voice of the author to rush to its aid, explaining it, answering for it. But notwithstanding Plato’s own literacy, these were the assumptions of an oral culture brought to bear on the technology of writing. To the modern mind, which has been gradually trained in the habits and assumptions of literacy, this vulnerability simply does not call up the same discomfort, and by the twentieth century it seemed reasonable to ask: what if we assume that the text says exactly what it intends to say, that it need not defend itself against any sort of oral cross-examination?
I should acknowledge that I find something deeply intuitive in all this. An important chapter of my own education occurred at St. John’s College, which is known primarily for its Great Books program. But just as important as its curriculum is how seminars are conducted. We would restrict our conversation to what the whole class can be assumed to have in common, and that extends not much further than our shared reading: “Let the text speak for itself,” we might say, rather than trying to explain it by pointing to historical or biographical facts known to have surrounded its composition. This sort of ‘explaining,’ after all, often means little more than “explaining away”. Nietzsche’s alleged syphilis infection, for instance, simply isn’t the most illuminating lens through which to look at the more obscure passages in Thus Spoke Zarathustra. These sorts of deflections are regarded as the last refuge of the superficial reader, someone unable or unwilling to attend carefully to what is actually printed on the page before them. Reading rigorously means proceeding in observance of the doctrine of sola scriptura: the text is assumed to be self-contained, inviolate, and impervious to the world outside itself which might try to tell us how to understand it.8 This is what it means to trust a book enough to let it guide you, rather than the other way around.
Writing in Noēma last year, the philosopher David J. Gunkel argued that “LLMs may well signal the end of the author, but this isn’t a loss to be lamented.” Drawing on Barthes’ obituary for the modern figure of the ‘author,’ Gunkel applies its implications to the rise of algorithmically-generated texts, which he thinks represent the twenty-first century realization by technical means of approaches to textuality earlier developed in twentieth century literary theory.9 There’s no reason to fret over scenarios like the one I opened with because our discomfort is rooted in long-discredited assumptions which posit authorship as the source of textual meaning: the outmoded idea that the text is a thread which connects the reader to some authorial apparition whose intentions and biography might determine its legitimate interpretation is finally laid to rest by the proliferation of texts which have no author at all. “[I]n responding to the ways in which these systems challenge how humans access, interpret and convey knowledge,” Gunkel concludes, “linguists, philosophers and AI experts tend to simply reassert concepts of authorship and authority that have long since been undermined.” If Roland Barthes tolled the author’s death knell, LLMs have arrived to bury the corpse.
The authorless texts which LLMs produce allow language to speak, at long last, for itself. The fact that there’s no identity behind its composition, algorithmically-generated text is not therefore ‘meaningless’ as the standard intuition might lead us to suppose. If the author’s ‘presence’ supplies a text with its meaning, then the author’s absence would have to repeal it. But this isn’t what happens: the rise of LLMs may abrogate the human author, but it leaves the reader intact, perfectly capable of getting lost in, say, an AI-generated novel, one perhaps custom-made for this reader in this moment. An extinction event for the author sets the stage for the golden age of reading.
There was always a liberatory purpose at work in Barthes’ thesis, which sought to escape the interpretive closedness of author-oriented reading into an open horizon of play and indeterminacy. The ‘author,’ after all, was always a crypto-theological concept hiding behind and presiding over a piece of writing and wielding explanatory mastery over it. Properly understood, Barthes had contended, “writing is the destruction of every voice, of every point of origin”, and the author’s only true power over the text should be the power that a reader is willing to grant. The death of the author means the emancipation of the reader, with whom the task of interpretation, untethered to traditional sources of authority, now exclusively resides. LLMs “can be liberating”, Gunkel argues. “They free both writers and readers from the authoritarian control and influence of this thing we call the ‘author.’” With the author deposed, the text achieves its unity and coherence in collaboration with the imagination of a creative reader, who now becomes the producer, rather than a mere consumer of meaning.
“Cursed (although I curse myself) be the hands that formed you!” Mary Shelley, Frankenstein
But reports of the author’s death are greatly exaggerated. This way of looking at LLMs vastly overstates the degree of consensus which has been established around Barthes’ thesis. The warm embrace of French poststructuralism inside the academy has cooled, and whatever appeal the idea might once have had among academics, news of the author’s demise has still to reach society at large, where the demand to tie a work to its author remains as powerful as that to retrace a crime to its perpetrator.10 In the popular imagination, biography and artistry remain intertwined in mysterious and controversial ways—and if you aren’t sure, you can ask Woody Allen or Roman Polanski, or any number of other figures whose works came under new scrutiny following revelations about the character of their maker. If Roland Barthes dealt a powerful blow to the concept of the author, the wound has not proved fatal. Had the sense of an author’s responsibility for their work truly been extinguished since the appearance of Barthes’ essay, Han Kang would not have found herself addressing a Nobel committee which had just recognized her for her work, Salman Rushdie might still enjoy the sight of his right eye, and, as a matter of fact, I would have seen no need to cite David Gunkel or stipulate his authority by identifying him as a philosopher: the text of his essay alone would suffice to make his case for him. Indeed, that we continue to invoke the name of Roland Barthes every time we speak about “The Death of the Author” is its own sort of implicit rebuttal. There simply must be more to authorship than this.
The longer I’ve lingered over the scenario I opened with the more I’ve become convinced that its roots reach all the way down to the basic rupture introduced by the technology of writing itself, a rupture which becomes a kind of abolition with the appearance of algorithmically-generated text. The death of the author does little more than dramatize and extend the decisive feature of textuality as such: the estrangement of the reader from the writer. The alienation of the text from its author sentences the author to ‘death’ the moment readers decide to treat the text as the solitary locus of hermeneutic ‘life.’ If nothing else, writing is the technology which facilitates the construction of meaning in its maker’s absence. As long as we could say we knew a text had been produced by a human writer, the author remained, as it were, on life support—largely irrelevant for the purposes of interpretation, but still the source of agency which initiated its composition. Only to the extent that we could assume human authorship were we justified in borrowing the norms of oral communication and applying them to the written word: namely, our sense that the text, like speech, is a creative gesture of a particular person acting upon the world, even long after it had parted ways with its author. This is precisely the assumption that LLMs overturn.
Every craft bears an ambiguous relationship to its maker: the made object is independent and autonomous, yet it remains in some sense morally tied to the personality responsible for creating it.11 An account of what we might call the ethics of authorship, then, would have to concern itself with the complexities of the ambiguous responsibility which authors have for their words, even after the two have become separated.12 It has always been obvious that to speak is also to act: every utterance is performative in the sense that words have power, however slight, to alter something about reality. Writing does the same: beyond its more quotidian uses, the written work plays a social and historical function, sometimes even articulating the self-understanding of an entire culture, which is then absorbed, taken up, and challenged by subsequent writers. Approaches which are only interested in the semiotic relationship of a text to other texts minimize or deny its origin in the agency of a particular, historically-situated writer. But all writing—all art—remains a moral endeavor if only to the extent that its consequences may echo indefinitely after the act of its creation has ceased. One of the great paradoxes of language is that, in contrast with the tangible applications of the hand and body, words themselves don’t work upon the world directly, yet they are the necessary means by which any real change in the contours of culture or history ever occurs; the very insubstantiality of the word conceals its power to act not from without, but from within those whom it has managed to touch.
To hear speech or to read text is also to ask “who is speaking?”, or, “who is writing?” The fact that a text is churned out by an LLM doesn’t release us from the question of its author’s identity; it just means that the ‘author’ is no longer a person, a ‘who’ in the traditional sense, but a corporate entity, both in the collective and in the capitalistic meaning of the word. LLMs don’t signal the death of the author so much as the author’s substitution by machines which by themselves cannot be the source of any initiatory action; this is why LLMs don’t invalidate the necessity of tracing a text back to its origins, but they do obscure its origins behind the composite automaticity of algorithmic processes: readers are not, as Roland Barthes had suggested, liberated to hear the voice of “language itself”: they are exposed to a textual confection, a novel sort of synthetic industrial product calibrated by interests that are much murkier than the more straightforward partnership between a writer and their publisher.13
The moment we shrug off the question “who is writing?”, separating the written word from whoever or whatever has produced it, literature takes definitive leave of its origin in the oral traditions of storytelling and rhetoric, traditions which were always rooted in and motivated by the sorts of moral concerns which arise as a consequence of our culturally- and historically-embedded situation. Over the full scope of their history, the two manifestations of the word—speech and writing—have proven to be less interested in the fabrication of an aestheticized ‘other’ world than in bringing some measure of articulated illumination to this one. In orality, the bard is the primary educational resource of the culture, and in literacy, the author has often played a similar role, uncovering the truth of the world and recording that truth for a future reader. The rise of language machines means not just the disappearance of the problematic figure of the author, but with it the withdrawal of the pedagogical function which real human authors have always played in the world, at whatever distance of time and space from their readers.
Some will see no more reason to lament the death of the (human) author than the death of the (human) computer, but the obvious difference is that we don’t live inside mathematics in the same way we do language. The reason I plan to keep writing is not that I believe no machine could ever produce a convincing imitation of my work, but because the labor of articulation is too important to a dedicated encounter with reality to ever condone its outsourcing.14 After the death of the author, Roland Barthes had wanted readers to become producers of meaning; the irony is that a world rich in texts yet poor in the human capacity to write is impoverished precisely by the fact that it has transformed the written word into yet another opportunity for consumption.
Longtime readers will know that I consider objections which take the form: “no machine could ever do X”, where ‘X’ is a capacity which human beings possess, to be naïve and shortsighted (see this essay from last spring). The structure of this argument reiterates the much older concept of the ‘God of the gaps’, now as a kind of ‘humanity of the gaps’. But capacities are precisely what machines can and do mimic. To define the ‘human difference’ in terms of our capacities means that our sense of our own humanity is never safe from the next technical breakthrough.
I’ve been jokingly referring to this account of textuality and authorship as “string theory.”
See Walter J. Ong, “The Writer’s Audience Is Always a Fiction”.
Or, as one commentator would later put it, “once woven, the web has no need of a spider.” (Seán Burke, The Ethics of Writing: Authorship and Legacy in Plato and Nietzsche, pg. 20.)
I have often thought that the College might adapt the motto of twentieth century phenomenology and declare: “Back to the texts themselves!” I’m curious how all this lands with Zena Hitz.
Nik Prassas recently coined the wonderful term “The Différance Engine”, a combination of Jacques Derrida’s concept différance with Charles Babbage’s “difference engine” from the very earliest days of computing.
Borrowing this from Seán Burke, The Ethics of Writing: Authorship and Legacy in Plato and Nietzsche, pg. 21.
Here, I use the word “morally” in the broadest sense to refer to the practical relationship between the text—and by extension the author—and the actual world.
Writing for The Atlantic last month, Deb Roy called LLMs’ separation of words from any agent who can be held responsible for them “a shift in the moral structure of language”. “An LLM’s words shape our beliefs, decisions, and actions,” Roy observes, “yet no speaker stands behind them.”
Last fall I wrote about the ways in which AI is both historically and structurally continuous with modern states and corporations.
“The labor of articulation” is a most helpful phrase which I borrow from L. M. Sacasas.
Very interesting article! Brings a few thoughts to mind:
It reminds me of debates I had in my religious days about the authorship of the Bible. So much of the religious tradition rested on the fact that it had divine authorship, and yet many wished to argue that the text itself was divine regardless of where it came from, because we decided to treat it as such. But given that Torah Study in particular was seen as a way to become more god-like in nature, and you take on the mind of the author you read by coming into relationship, I wonder what sort of being you take on when you enter the 'mind' of an authorless LLM, and what that will do to readers who take it seriously. It's hard to imagine something religion-like not springing up.
Also, the idea of custom novels feels a bit off-putting too. There is something that feels wrong about being able to change the ending of a book you do not like, for example, or have a book written just for you. On the one hand, that power would be incredible to use, but also I think reading is supposed to be an encounter with limits. The text stands on its own, and changes you, and forces you to wrestle with it, with something that stands outside of yourself. Being able to control or change it on command disrupts our entire sense of a shared reality, in addition to endorsing a sort of narcissism that prevents true formation from occurring. Almost like an algorithmic social media filter bubble on steroids. I think the accelerated collapse of an agreed upon shared reality (which is already happening) is also a huge problem when you lose the authority of authorship.
In any case, it feels nice that I can engage with you (a human being who I could enter into relationship with) and not a faceless pile of bits here!
Very tangential, but it makes a lot of sense that you're a Johnny. Often makes me regret not going there when I meet Johnnies and see how much fun they are in conversation.