AI as a Fiction Machine: The Limits and Potential of Large Language Models
Modern artificial intelligence (AI) has achieved remarkable fluency in language, enabling conversational fluidity akin to human interaction. However, a crucial understanding is often overlooked: Large Language Models (LLMs) aren’t designed for truthfulness, but rather to generate narratives that “make sense” within a given context. This perspective, championed by machine learning researcher Léon Bottou, frames LLMs as “fiction machines,” prioritizing coherence over factual accuracy.1
How LLMs Construct Narratives
LLMs operate by predicting what should come next in a developing narrative. They are trained to generate text that aligns with general structures of language, allowing them to transfer these structures to new and unfamiliar situations. This ability relies on principles like “compositionality,” where the meaning of a complex expression is derived from the meanings of its parts and their combination.1 LLMs can readily imagine and create ideas not explicitly present in their training data, even when the context is entirely novel.
Despite not being explicitly programmed for accuracy, LLMs often exhibit surprising truthfulness. This may be attributed to intensive reinforcement learning from human feedback (RLHF), where human validators fine-tune responses to be correct or socially acceptable.1
Can AI Write Novels?
Given their capacity for narrative generation, the question arises: can AI create novels? Bottou suggests that creating new plots should be relatively easy for LLMs, as fiction machines are inherently equipped to construct stories, regardless of their quality.1 The AI essentially “prints fiction on a tape,” borrowing facts from its training data and filling gaps with plausible fabrications.
The Challenge of Novel Theories
However, generating entirely new theories presents a greater challenge. While LLMs can easily identify the correct model from a set of candidates, formulating a theory requiring new concepts is far more complex. This necessitates either assigning new meanings to existing words or creating entirely new concepts – a significant hurdle for current AI capabilities.1
Historical scientific breakthroughs, such as Einstein’s theory of relativity or the development of thermodynamics and quantum mechanics, involved redefining existing terms or introducing entirely new concepts (like photon, quark, and entropy).1
The Limits of Symbolic Representation
theories typically require a causal structure and mathematical formulation, demanding that phenomena be understandable to humans through symbolic representation. This raises a fundamental question: can intelligence be entirely specified in terms of symbols? If aspects of intelligence, such as emotion, visual imagery, and motor control, aren’t fully captured by symbols, it may be impossible for humans to comprehend a novel theory generated by AI if the machine cannot explain it using understandable symbols.1
This concept echoes Geoff Hinton’s metaphor of AI as an “alien” – an intelligent entity created by us, yet potentially thinking in ways fundamentally different from our own, making complete understanding a challenge.1
LLMs in Narrative Analysis
Recent research demonstrates the potential of LLMs, specifically Claude and GPT-o1, as tools to assist in qualitative research, particularly in narrative analysis. A study comparing LLM-conducted analysis with human analysis of 138 stories found that LLMs could conduct thorough and credible narrative analysis, with findings comparable to those of human researchers.2 This suggests LLMs can be valuable assets in analyzing qualitative data, though researchers must consider how to best integrate these tools into their workflows.2
Open-Source Tools for AI Storytelling
The rise of open-source tools like AI StorySmith further democratizes AI-powered storytelling. AI StorySmith allows users to generate entire books using local LLMs (like Dolphin 2.6 or Mistral) without relying on cloud-based services, prioritizing privacy and offline creation.3
As LLMs continue to evolve, understanding their inherent nature as “fiction machines” is crucial for harnessing their potential while acknowledging their limitations. The future may hold intelligent entities that think differently than us, requiring us to learn their “language” to fully comprehend their contributions.