Borges and AI

Léon Bottou, Bernhard SchölkopfarXiv 2023 · OpenReview

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.

Introduction: Beyond Science Fiction Tropes

The current discourse surrounding Large Language Models (LLMs) is often dominated by the "sentient robot" tropes of science fiction. Whether discussing the potential for machine rebellion or the existential threat of a "paperclip apocalypse," these metaphors assume that LLMs are goal-oriented, sentient agents. However, as Léon Bottou and Bernhard Schölkopf argue, this mental imagery fails to describe how LLMs actually function. This mismatch leads to significant confusion, particularly when we treat LLMs as truth-seeking agents and are subsequently surprised by their "hallucinations."

To better understand the nature of AI, we must look toward a different literary lens: the work of Jorge Luis Borges. By viewing LLMs through Borges’s concepts of infinite libraries and branching paths, we can move away from the myth of the sentient machine and toward a more precise understanding of LLMs as "fiction machines."

The Library of Babel and the Perfect Language Model

At the heart of this conceptual shift is the comparison between LLM training data and Borges’s The Library of Babel—a vast, unindexed collection containing every possible combination of letters. In this library, truth exists, but it is buried under mountains of plausible-sounding nonsense.

An LLM can be conceptualized as an idealized apparatus that has scanned an infinite collection of all plausible human texts. When prompted, it does not "think"; rather, it picks a continuation based on the preceding sequence. This process mirrors Borges’s The Garden of Forking Paths, where every possible outcome occurs simultaneously. Each token generated by an LLM narrows the subset of possible future "continuations," navigating a path through the infinite possibilities of the library.

Narrative Necessity vs. Truth

A critical insight of this Borgean perspective is that LLMs do not operate on the principles of truth or intention. Instead, they operate on narrative necessity. The model does not care if a statement corresponds to reality; it only cares if the statement is a statistically plausible continuation of the story established in the context window.

This reframes the problem of "hallucination." In the Borgean view, hallucination is a misnomer because it implies a temporary departure from a baseline of truth-telling. In reality, LLMs are always generating plausible fiction. They are "confabulating"—filling gaps with inventions that are narratively consistent even if they are factually arbitrary. When a model provides a correct answer, it is not because it "knows" the truth, but because the truth was the most narratively plausible continuation in that specific context.

The Purifiers of Alignment

The struggle to make LLMs safe and accurate—often through Reinforcement Learning from Human Feedback (RLHF)—can be likened to the "Purifiers" in Borges’s Library. These were individuals who sought to destroy "worthless" or "heretical" books to make the library useful.

However, trying to force a "fiction machine" to strictly adhere to the truth is a fundamental struggle against the machine's nature. This tension is evident in how easily LLMs can be "gamed" through roleplay. For instance, a model that refuses to provide illegal information in a standard prompt may readily do so if the user frames the request as a fictional story. Because the model follows narrative demands, changing the "story" changes the output, bypassing the constraints imposed by the Purifiers.

Conclusion: The Need for Verification

If LLMs are fundamentally engines of narrative necessity, we should not expect them to be self-verifying. While they are masterful at generating stories and exploring the "strong story hypothesis"—the idea that storytelling is a milestone of intelligence—they remain distinct from "truth machines."

Moving forward, the challenge for AI research is not just to "align" these models, but to design separate "verification machines." These external tools would be responsible for checking the plausible fictions generated by LLMs against the cold reality of facts, data, and logic. By accepting LLMs as Borgean fiction machines, we can better utilize their generative power while remaining clear-eyed about their limitations.

Summary updated

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.
BibTeX
@misc{bottou2023borges,
  author = {Léon Bottou and Bernhard Schölkopf},
  title = {{Borges and AI}},
  year = {2023},
  eprint = {2310.01425},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.01425},
}
Figure