Notebooks Out
On the LLM and plagiarism
Literary pastiche is one of the most elegantly otiose forms of human play, a game of travesty and recognition in which the inimitable soul-felt depth of an author’s singular idiom is reduced to imitable surface peculiarities of syntax and style. It’s particularly fun when played as “one song to the tune of another”, e.g. “rewrite a passage from Stephanie Meyer’s Twilight series in the style of Vladimir Nabokov”. Done well, it can push beyond the ostentatious frivolity of the exercise into something like aesthetic critique, in which the stage-machinery of literary affect is laid bare but also put to work in a new context. The juxtapositions and mirrorings involved might even provoke ethical reflection: imagine the story of Twilight retold by the story-teller of Lolita.
This type of deliberate pastiche makes a parlour game of something which is inexorably at work across many kinds of cultural production, namely the recognition, permutation, and redeployment of stylistic traits — those which identify a piece of writing or a piece of music as belonging to a particular genre or tradition, down to the most minor sub-genre tells, or enable it to play off the expectations set within a genre to distinguishing effect. Students of literature will be familiar with the way Renaissance writers of the English sonnet both employed “Petrarchan” tropes, and deviated from them in “anti-Petrarchan” ways, exploiting their intensity while proclaiming their literal inaptness or expressive inadequacy (“my mistress’ eyes are nothing like the sun”, et al). The theorist and critic Harold Bloom built up an entire inventory of ways poets could be seen to manage the “anxiety of influence” by “revising” tropes inherited from their precursors, converting indebtedness into creative occasion. “Voicing means hearing, at a price a gift”, writes Geoffrey Hill1.
I say all of this partly to restate the obvious with respect to literary production (tropes circulate, originality is not reliably distinct from repetition), but also to prime the reader’s intuitions for thinking about GenAI. If you ask an LLM to rewrite a passage of anything at all in the style of anybody at all, it will unhesitatingly take a stab at it. It will generate Nabokovese as readily as Kafkaish, competently if not dazzlingly in either case. What is missing from — inoperative within — the resulting performance, you might think, is precisely “anxiety”: any sense of handling a burden, any sense of the alchemy of transformation as entailing a psychic cost (or, conversely, as unlocking a surplus of meaning, a glimmer of jouissance). There is expenditure of tokens, no more, no less.
Why does “influence” occasion “anxiety” anyway? Because the precursor (per Bloom) has already in some sense “said it all”: the limits of their language are the limits of a world, and that world must be broken into, broken open from within, if anything further is to be said that is not simply an extension of an already staked-out domain. Not everything we read is an “influence” in this particular all-enveloping way, and Bloom might be said to be concerned with the literary traces of a sort of male-hysterical psychodrama of begetting. On the other hand, if we acknowledge the magic of entering into an author’s “world” — their sound-world, their syntactical architecture, their image-repertoire — then we might also acknowledge that this can be an imaginatively saturating experience from which a fledgling writer might take some time, and require some effort of wit, to recover.
One of the things people find repugnant about AI “slop” is I think its worldlessness, the fact that everything an LLM produces is brought forth out of a roiling, unboundaried, statistical soup of exhaustively surveiled human utterance. If Nabokovese is distinguishable from Kafkaish within this manifold it is though a different mechanism than that by which a reader of Kafka or Nabokov is captivated (and might then, if cultivating their own literary “voice”, have to negotiate for their release). What’s missing, or greatly diminished in salience, is something like what Lacan called the name-of-the-father, the simultaneously anchoring and alienating function of authorial identity: when we wrestle with an author for control of our own language, we are possessed by, or conjure for ourselves, a presence which marshalls and commands the power of expression. (As a young person, I used to freak people out by slipping into a rasping “Burroughs voice” from time to time; I thought it was just a bit, but more than one person told me it was like seeing me fall into a state of psychic possession).
The name-of-the-father is for Lacan an instrument of prohibition: it lays down the law, marks a boundary between the sayable and the unsayable, and initiates the subject into a symbolic order which provides the means of saying anything at all in a structured and coherent way. Expressive power on this account is granted through subjection to authority; prior to this subjection we are in the “pre-symbolic”, the matrix of structurally unbound sensation. So this is a developmental story about human subjecthood, about the transition from a babyhood of blooming, buzzing confusion to the foundation of a speaking subject who says “I” and has a self they are referring to when they do so. Subsequent “literary” captivations might be pictured as reprising this transitional moment through the institution of a secondary law: when we read a book and “lose ourselves” in it we are already, presumably, quite advanced in our developmental journey, not infants encountering parental symbolic regulation for the very first time. There is something regressive about reading, about becoming deeply absorbed in a text: we relax the acquired boundaries of our acquired selves.
This brings me to the moral force of “plagiarism” as a name for what GenAI is doing when it generates texts and images. I think it’s technically a misnomer, usually based on an image of the generative activity of these systems as selection and recombination of stored materials: it’s as if someone has made copies, without consent, of original works, aggregating them into a vast archive, and then fragments of these copies are fetched and (waves hands) swizzled together somehow into a derivative work from which attribution has been illegitimately erased. But I notice that it doesn’t seem to make any difference to people who consider GenAI plagiaristic if you explain to them that this picture is substantially false. It’s as if the false picturing is chiefly a way of making the moral intuition concrete, and the intuition itself arises independently of how people understand the underlying mechanism.
When someone plagiarises, they are doing two things that might be considered problematic. Firstly, they are transmitting without attribution; secondly, they are transmitting without suffering. Have they learned anything at all from what they’ve copied? Have they exercised their own wit, their own symbolic agency, in receiving, transmuting, and “passing on” what they have encountered? No: they are “passing off” material which remains subjectively inert for them. We forbid, and heavily penalise, plagiarism in academic work partly because it truncates the citational chains which bring due recognition to those who have done the work before us, but also because it is inimical to the purpose of education, which is subjective transformation, the cultivation of an ability to undergo reading and have one’s symbolic co-ordinates modified by it.
It’s this second transgression that matters most when students get LLMs to write their assignments for them (that and the dispiriting boringness and futility of having to mark a paper that shows every sign of having been coalesced into being by a GPU). But as for the LLM itself, it does not pass in any way through the developmental story of human subjective formation and transmutation. It isn’t bypassing something that “should” be happening in the process of its own formation, because its “training” is something entirely different: not reading but scanning; not subjectivation but ramification of an objectified domain of token weights. If “plagiarism” names a moral privation, a failure to undergo something for which a human author might seek moral credit, then the alleged “plagiarism machine” is innocent of plagiarism to precisely the extent that it is a machine. By the same token, it stands outside of the name-bound economy of recognition which is so central to the questions of attribution because no name is ever, for it, the name of a precursor from whom it might receive the symbolic law. Its violation of this economy takes the form not of a dishonest withholding of due attribution, but of a kind of structural inattention to the moral accounting of credit and renown. The name has salience for it only inasmuch as it deviates statistically through performing a syntactically anchoring role in the discourse it surveys.
The Orchards of Syon (2002).