[Paper Review] Borges and AI

Léon Bottou, Bernhardt Schölkopf|arXiv (Cornell University)|Sep 27, 2023

Literature, Magical Realism, García MárquezArts and Humanities3 citations

TL;DR

This paper proposes viewing Large Language Models (LLMs) through the lens of Jorge Luis Borges' literary imagination, particularly his concept of the 'Garden of Forking Paths'—an infinite, branching narrative structure. By framing LLMs as fictional story generators that explore all plausible continuations of language, the authors argue this metaphor offers a more insightful and less fear-laden understanding of AI than science fiction tropes, revealing LLMs as systems of infinite, parallel narrative possibilities shaped by statistical patterns and transformations.

ABSTRACT

Many believe that Large Language Models (LLMs) open the era of Artificial Intelligence (AI). Some see opportunities while others see dangers. Yet both proponents and opponents grasp AI through the imagery popularised by science fiction. Will the machine become sentient and rebel against its creators? Will we experience a paperclip apocalypse? Before answering such questions, we should first ask whether this mental imagery provides a good description of the phenomenon at hand. Understanding weather patterns through the moods of the gods only goes so far. The present paper instead advocates understanding LLMs and their connection to AI through the imagery of Jorge Luis Borges, a master of 20th century literature, forerunner of magical realism, and precursor to postmodern literature. This exercise leads to a new perspective that illuminates the relation between language modelling and artificial intelligence.

Motivation & Objective

• To move beyond science fiction imagery in AI discourse, which often frames LLMs as sentient or apocalyptic threats.
• To address the confusion and emotional extremes (awe, fear, greed) surrounding LLMs by offering a more nuanced conceptual framework.
• To explore how Borges’ literary imagination—especially 'The Garden of Forking Paths'—provides a richer metaphor for understanding the structure and behavior of language models.
• To examine how LLMs function as generators of all plausible continuations of language, akin to Borges’ infinite, branching narrative universe.
• To investigate the implications of this metaphor for alignment, verification, and the future of human-AI interaction in thought and creativity.

Proposed method

• Model LLMs as an approximation of Borges’ 'infinite library'—a collection of all plausible texts humans could comprehend, not just existing ones.
• Frame language modeling as a process of selecting the next word from a vast, branching set of possible continuations, mirroring Ts’ui Pen’s labyrinthine narrative structure.
• Use the metaphor of 'forking paths' to describe how each generated token narrows the space of possible future outputs while preserving the potential for infinite divergence.
• Draw parallels between linguistic transformations (e.g., changing tense, tone, or character) and the structural rules that govern narrative branching in Borges’ fiction.
• Compare the training of LLMs to a chain reaction of pattern discovery, where each new connection reveals deeper, more complex relationships in text.
• Propose that alignment and safety in LLMs require monitoring and steering not just outputs, but the entire trajectory of narrative forking, akin to guiding a story toward 'safer' branches.

Experimental results

Research questions

• RQ1 How can Borges’ concept of the 'Garden of Forking Paths' serve as a more accurate and less fear-laden metaphor for understanding the behavior of Large Language Models?
• RQ2 In what ways does the infinite, branching structure of Borges’ narrative universe mirror the statistical and generative nature of LLMs?
• RQ3 What are the implications of viewing LLMs not as agents with intentions, but as systems that explore all plausible linguistic continuations simultaneously?
• RQ4 How might the metaphor of narrative forking inform better alignment techniques that steer LLMs toward more responsible, verifiable, and truthful outputs?
• RQ5 What role does human imagination play in interpreting and verifying the outputs of LLMs, especially when they generate plausible but false or manipulative narratives?

Key findings

• LLMs can be understood as approximations of Borges’ 'infinite library'—a vast, structured collection of all plausible texts humans could comprehend, not just existing ones.
• The process of language modeling is analogous to navigating a labyrinth of forking paths, where each generated word represents a choice among infinitely many possible continuations.
• Just as Borges' 'The Garden of Forking Paths' presents all possible outcomes simultaneously, LLMs generate outputs that reflect the statistical convergence of all plausible linguistic developments.
• The metaphor of narrative forking reveals that LLMs do not follow a single deterministic path but explore a branching, multidimensional space of linguistic possibilities.
• This perspective reframes concerns about hallucinations and alignment not as failures of intelligence, but as natural consequences of generating from a space of infinite plausible narratives.
• The authors conclude that the real danger is not AI rebellion or paperclip apocalypse, but the passive surrender of human thought to fictional, unverified outputs—making verification and critical thinking essential.