Large Language Models and the Death of the Author

In my last Slaw column, I dealt with the rapid responses to the “authorship” question from the leading journal Nature and the U.S. Copyright Office to the sudden arrival of large language models (LLM), such as ChatGPT. Both publisher and government agency made it clear that they will not accept such works for publication or copyright. More recently, Nature reported this June that it will now require authors to state that their submission does not use AI-generated images.

With the state and impact of LLM continuing to rapidly evolve, I want to follow with further reflections on the authorship and thus intellectual property implications. In particular, I’ve become struck by how LLM are demonstrating, in effect, some of the more radical concepts of authorship advanced over a half-century ago by French postmodernists, such as Roland Barthes, whom I’ll deal with here, but also Michel Foucault (for another time). Given how hard it can be to get in on the learning behind machine learning – with its traceless neural networks processing massive data sets – we must seek out lessons where we can find them.

Now that authorless texts are a thing, it may call to mind, at least for some of us, Barthes’ essay The Death of the Author, published during those heady Parisian days of May 1968. In it, Barthes, eager to historicize cultural artifacts, goes after the modern tendency to venerate authorship in a sentimental vein. He cheekily calls on celebrated poets, some of whom are too ready to attest to authorship’s exaggeration. In the nineteenth-century poet Stéphane Mallarmé, Barthes finds a champion for how “it is language which speaks, not the author.” The later poet Paul Valéry directed our attention, Barthes insists, to “the verbal condition of literature,” while regarding “recourse to the writer’s interiority” as “superstition.” This deflating (if not slaying) of authorship makes of the text, in Barthes words, “the pure gesture of inscription (and not of expression)” that “traces a field without origin – or which, at least, has language, which ceaselessly calls into question all origins.” There is, then, only the “variety of writings, none of them original, [which] blend and clash” in “a tissue of quotations drawn from the innumerable centers of culture.”

Barthes’ description strikes me as an apt characterization of ChatGPT. This LLM responds to a prompt with gestures of pure inscription across from the innumerable centers of culture. Yet seeing connections between ChatGPT and the theorizing of a man who favored the fountain pen seems a little far-fetched, you need not take my word for it. I asked ChatGPT about the relation of its work to Barthes’ essay. It did not hesitate to confirm that “the texts produced by the [ChaptGPT] model can be seen as an embodiment of Barthes’ ideas… as it separates the text from a specific author’s authority and places the interpretive power in the hands of the reader.”

This micro-display of intellectual bravado (well, compared to “acing” the bar exams) did, however, manage to overlook Barthes’ intellectual property critique of this “modern figure” of the author, symbolizing “the prestige of the individual.” It is nothing less than the “culmination of capitalist ideology,” for Barthes, “which has attached the greatest importance to the ‘person’ of the author.” I have seen this confirmed with the origins of copyright in the early eighteenth-century. To great effect, the Stationers’ Company employed the revered author as a front in promoting publisher property rights – with references to “Many Learned Men [who] have been at great Pains and Expence in Composing and Writing of Books” – in successfully lobbying for the passage of the Statute of Anne 1710, considered to be the first modern copyright law.

The celebration of authors – as per my Barthean tissue of quotations above – is more, I have to believe, than simply capitalist ideology at work. But it is also that. And decidedly so, in my and others’ efforts to bring open access to research publications. After all, scholarly publishers hold that their profits are in the service of their authors. Yet researchers, as both their authors and readers, would be far better served by sustainable universal open access (as would scientific progress and thus humankind). This could be facilitated, I have proposed, by copyright reform that introduces statutory licensing for research publications. We could set aside the author’s (which is to say the publisher’s) capitalist right to exclusive access in favor of open research publications, for which universities and funders could fairly compensate – at rates set by the Copyright Board – their publishers.

Yet I am also beginning to think that we have something of an existential nature to consider in this encounter of 1960s Paris and 2020s San Francisco. In explaining to me its Barthean state, ChatGPT refers to how “as an AI language model, ChatGPT does not possess a personal identity, intentions, or consciousness. It doesn’t have an author in the traditional sense. Instead, it processes and generates text based on patterns and information it has been trained on.”

Barthes would object that we are all, in effect, learning to process and generate texts “based on patterns and information [we have] been trained on.” And as part of that learning, we glean patterns in ourselves and others that we speak of in terms of personal identity, intentions, and consciousness. We do not so much “possess” these qualities (within that “interiority” referred to above), as read them off of the generated texts. Barthes concludes his essay with the force of the reader “that someone who holds together in a single field all the traces by which the written text is constituted.” This I offer tentatively, as a reader of uncanny chats.