Are algorithms and LLMs changing our conception of literature

Computerized large language models (LLMs) are making inroads into the realm of literature. Their ability to generate coherent texts and mimic all manner of writing styles has sparked lively debate among writers, literary theorists and researchers. Some fear a threat to human creativity while others think it's worth exploring the innovative potential of LLMs.

Marcello Vitali-Rosati, a professor in Université de Montréal's Department of French-language Literatures, sees both creative promise and built-in limitations to these tools. In his research, he considers whether LLMs can transform the very idea of literature and examines the social, economic and cultural issues they raise.

A new era of experimentation

Large-scale language models provide fertile ground for writers who want to test the limits of language and machine. Some authors are wielding them subversively to expose the flaws, limitations and grey zones of artificial intelligence, rather than simply using them as productivity tools.

Vitali-Rosati believes these inventive ploys can open up new paths for contemporary literature.

"Authors are playing with the algorithms to probe new creative byways," he said. "For example, some are trying to exhaust the limits of the algorithms," questioning the machine itself, driving it hard to see how far it can go before it stumbles.

One such writer is French playwright and poet Milène Tournier. First she asked OpenAI's ChatGPT for 27 definitions of "muraille de Chine," which can mean a Chinese wall or the Great Wall of China. Then she asked the algorithm to shorten the definitions, and to keep shortening them.

"In the end, only one definition remained: 'wall.' It demonstrates the dilemmas of the algorithm and the limitations of language itself," said Vitali-Rosati. His stratagem exposed the boundaries of the model, putting into question the very nature of definition and linguistic reduction.

Experiments with LLMs are not confined to attempts to wear down the algorithm. Other authors are exploring co-creation with AI, using the models' suggestions to generate ideas or text fragments that might not have come to them otherwise. This collaborative process can yield hybrid works produced through dynamic interaction between human and machine.

Towards standardized language and thinking?

While LLMs can open up new channels of creativity, they also raise red flags about standardized language and conformist thinking.

The algorithms are designed to produce texts that can pass coherence and relevance tests based on statistical criteria drawn from huge corpora of textual data. This approach tends to produce fairly homogeneous texts. As such, it can run counter to the very essence of literature, which often seeks to subvert norms and roam the margins.

"What's problematic is the slickness of these algorithms," Vitali-Rosati said. "They are mainly designed for utility, to satisfy the demands of a worldview rooted in functionalism and productivity." This functionalism can lead to a brand of standardization that steers LLM-generated texts away from the depth, complexity and rough edges that are the hallmarks of literature.

Programmed to efficiently meet user expectations, LLMs can produce texts that are pleasing and well-structured but lack what literature often seeks to explore: the dysfunctional, the unknown, the peculiar. "Literature looks for such points of friction," said Vitali-Rosati, "and the difference between texts of literary merit and texts of less aesthetic interest is precisely the ability to find these shadowlands."

Continuity or a break with the past?

Are LLMs changing our concept of the literary? According to Vitali-Rosati, it's complicated.

He doesn't see LLMs as a radical break that will transform the nature of literature but rather as a continuation of existing practices. "I don't think LLMs are revolutionary per se; they belong to a much longer tradition," he said. In his 1941 short story "The Library of Babel," for instance, Jorge Luis Borges imagined a world in which literature is created through a quasi-algorithmic process.

Vitali-Rosati also pointed to literary practices such as plagiarism, pastiche and retelling that have always been part of the literary landscape. He views LLMs simply as an extension of these practices, one that makes them more visible and accessible through technology. It doesn't mean literature is losing its essence, he said, but rather that it continues to evolve by incorporating new tools and ways of writing.

'An impact on what we can write'

"Any technical tool we use to write has a major impact on what we can write and what we can think," said Vitali-Rosati.

He cited the example of the French novelist Gustave Flaubert, who spent much time sharpening his quills with knife-edge precision. He chose his quills himself and said he couldn't write unless they were perfectly trimmed to his requirements.

Similarly, Paul Valéry used an elaborate system of methodically arranged colour-coded cards to prepare his texts. He had notebooks of specific sizes and different pens, and the physical arrangement of his materials influenced his thinking and writing processes.

"The grain of the paper, the size of the sheet, the type of pen, the colour of the ink: these details, which are often considered trivial, are not so at all," said Vitali-Rosati. "They are the stuff of writing and they leave their stamp on literary production."

Vitali-Rosati situates large language models that can generate massive reams of text within the broader framework of the technical, economic, cultural and historical influences that shape literature.

He believes that Microsoft Word has had an even more profound, though often overlooked, impact on literature.

"Whereas every writer used to have their own style of notebook, type of pen, and preferred ink and paper, today everyone is in the same digital environment," he pointed out. "This conditions their writing in telling and unfortunate ways."

This standardization of writing tools bends literary practices to the norms of desktop software, squeezing writing into formats designed for document management rather than literary creativity.

"I don't think LLMs will have a real impact on literature," Vitali-Rosati concluded. The literary marketplace will change, perhaps even the way we write will evolve. I'm not saying that won't have an effect, but it won't be of a different order than the impact of any other technological device used for writing.

"To be sure, writing with a large language model will alter the way you write - just as a spell checker, Word, or the way you sharpen your quill would."