This article was automatically translated from the original Turkish version.
The idea that the author holds primary authority over the text and speaks from within himself began to be challenged in the 20th century. In response to this shift, Barthes metaphorically proclaimed the death of the author and announced the advent of a new mode of textual production. For Barthes, the author is merely an intermediary who transmits to the reader a vast lexicon of language and culture—a “tissue of citations.” The author no longer possesses the authority attributed to him in classical author-centered texts. The true meaning of a text emerges only during the act of reading and solely through the reader’s interpretation. From the moment a text is written, it becomes independent of its author and opens itself to infinite interpretive possibilities. In this sense, precisely as Barthes metaphorically stated, the author is dead.
Barthes grounded this bold argument in the emergence of the author as a subject within modern, individualized societies shaped by modernity and capitalism. Since the essay was written in 1967, technological and sociological developments have, to a significant extent, confirmed this thesis. The author and authorship, once a crucial subject in early modern society, have seen their influence constrained by the proliferation and democratization of publishing opportunities, the rise of digital publishing, and the liberating nature of the internet. Now, the authority of the text and the scope of its interpretation no longer reside within the author’s domain but expand according to the reader’s chosen context. The text has been fully severed from its author and handed over entirely to the reader. The author who once designed the entire text has vanished. Today, the advent of AI-generated authorship opens new dimensions to this debate. Yet it is inevitable that contemporary debates surrounding AI will not resolve the lingering question of the author’s death.
How Do Language Models Model?
Before addressing the relationship between AI models and text or authorship, it is essential to understand how language models operate. Rather than memorizing words (“tokens”), these systems advance by predicting the next word based on statistical probabilities derived from linguistic patterns learned across billions of parameters. Large language models (LLMs) are typically trained in two main stages: First, they undergo pre-training on vast and diverse datasets to learn general linguistic structures and grammar. In this phase, their primary task is to predict the most probable next word given a context. Then, through post-training or fine-tuning processes involving human feedback, they are refined to produce outputs that are more consistent, safe, and aligned with user instructions. Thus, the entire process can be characterized as the organization of word sequences generated by statistical outcomes.
Although many linguists and even philosophers have recently argued that LLMs truly “think,” a conceptual crisis of ambiguity persists. It is impossible to claim that these systems, which merely calculate sequences of words, comprehend words or grasp their meanings. The key to recognizing this impossibility may lie in the fact that human processes of understanding and comprehension cannot be reduced to such statistics. Since machines cannot experience the meaning inherent in things—for instance, when we say “red,” we recall the color red and the lived experience of redness, which machines cannot replicate—it seems implausible to call the statistical operations of LLMs acts of understanding or comprehension.
Can Statistics Be Authorship?
It is well known that the rapid development of AI has sparked debates across many fields. The most publicly visible of these debates is the legal conflict between copyright holders and AI companies, naturally extending into theoretical dimensions. One of the most prominent recent developments emerged from a study conducted last month at Cornell University, which found that Meta’s LLaMA language model, version 3.1, could reproduce up to 42 percent of J.K. Rowling’s “Harry Potter and the Philosopher’s Stone” word-for-word.to be able to produce The study compared various LLMs in their ability to regenerate popular books and claimed that LLaMA 3.1 had memorized certain texts—including “Harry Potter and the Philosopher’s Stone” and “1984”—from its training data. It was shown that the model could accurately regenerate specific passages of these books in 50-token sequences, despite never having been explicitly exposed to them during training.
Researchers explain this “memorization” by suggesting possible errors during training, repeated exposure to the same dataset, or the model having encountered excerpts from the books on other websites. A similar issue arose in the 2023 copyright lawsuit filed by The New York Times against OpenAI and Microsoft. In defending itself against similar accusations, OpenAI argued that its models provide “transformative” use, do not “copy” original works, and generate novel word combinations as a result of their learning processes.defended However, LLaMA 3.1’s ability to regenerate Harry Potter at such high fidelity may challenge this “transformative” argument legally; while there are differing views on whether copyrighted books should be used to train AI, such studies reinforce the perception that authors and publishers’ works have been copied without permission.
Leaving legal proceedings aside, if we return to Roland Barthes’s essay in the context of authorship, a model like LLaMA may effectively complete the erosion of the already weakened authorial subject—the very function of textual production (scriptor) that defines our current era. As previously noted, contemporary authorship, limited to arranging a specific “tissue of citations,” can be displaced by AI-based language models that construct texts by selecting and organizing from vastly larger, personalized datasets.
When LLaMA generates a fictional text resembling Harry Potter, it does so not by understanding magic or characters in human terms, but by systematically recombining patterns, grammar, and vocabulary learned from vast existing texts. This new process of text generation successfully advances the notion of authorship that, as Barthes indicated, is detached from human consciousness—already severed from the author’s personal experience, intention, and life story. Thus, while the text was detached from its author in the 20th century, in the 21st century it is now also detached from the very notion of being written. As a result, it becomes uncertain whether the speaker is truly an experiencer or merely a statistical construct.
The authorial subject in modern society, as Barthes described it—individualized and alienated from nature—had already lost some of its authority over the text after postmodernism. With the integration of AI technologies into text production, authors whose domain of textual authorship is shrinking may turn toward pre-modern, performance-based oral literary forms such as storytelling, epic narration, or meddahlık—or toward entirely new narrative forms we have not yet witnessed. However, the full realization of these alternatives depends on the direction and extent of AI development, the policies of states and corporations, and the evolution of legal frameworks and processes.
In conclusion, the future of this complex landscape remains uncertain. As AI technologies rapidly evolve, the tension between copyright holders and AI companies necessitates the continuous redefinition of the author’s literary and legal existence. Perhaps, within the framework of Barthes’s “Death of the Author,” we are not witnessing the complete extinction of creativity and authorship, but rather the emergence of new forms of collaboration between humans and machines—or, as Barthes himself concluded in the final sentence of his essay, “The birth of the reader must be at the cost of the death of the author.”*Written by Emirhan Kartal.