The ‘Fidgeting’ Tell: Why AI Still Struggles to Mimic Genuine Literary Prose
Table of Contents
The Ghost in the Machine
For years, the conversation around generative AI has focused on productivity: the ability to summarize a meeting, draft a legal brief, or write a functional piece of Python code. But a more existential question is beginning to emerge in the literary world: can an LLM actually produce prose that humans want to read, or is it simply mirroring the shape of a story without understanding the soul of it?
The debate reached a boiling point recently when Granta, a prestigious literary magazine, published a Commonwealth Prize-winning story by Jamir Nazir. Readers quickly flagged the piece for what they described as “AI hallmarks”—strange repetitions of specific words and metaphors that felt structurally sound but logically “constipated.” The ensuing controversy highlighted a growing tension in the creative arts: the gap between a machine that can mimic a style and a writer who can command a narrative.
Identifying the ‘Robot Tells’
Most users can spot rudimentary AI writing through obvious linguistic patterns. In the case of models like Claude or GPT-4, this often manifests as an over-reliance on em dashes, a penchant for “not X but Y” constructions, and the sudden, inexplicable appearance of the word “delve.” However, when AI attempts to emulate the masters—George Eliot, James Joyce, or Ernest Hemingway—the tells become more subtle and more psychological.
Experimental testing involving “vibe-coded” games—where users attempt to distinguish between public-domain literature from Project Gutenberg and AI recreations—reveals a persistent flaw: the AI’s tendency to create characters who do absolutely nothing.
In AI-generated scenes, characters often exhibit a peculiar, repetitive physicality. They fidget. They run a finger along the edge of a table; they adjust a collar; they stare blankly at a fireplace. These are not character-building details, but rather “filler behaviors” the model uses to simulate a scene when it doesn’t know how to drive a plot forward.
The Problem of Inertia
The fundamental issue is one of agency. A human author uses a description of a character looking at a fireplace to signal a specific emotional state or to pivot the scene’s tension. An AI uses the fireplace because it has seen thousands of descriptions of fireplaces in its training data.
Take, for example, a generated passage in the style of Henry Fielding. The characters may sound correct—the cadence of the 18th century is there—but the action is inert. The characters exist in a state of perpetual ambivalence. They look at each other with expressions that cannot be interpreted, and they speak in qualifiers (“as if,” “as though”). There is no certainty and, consequently, no narrative momentum.
Can the Model Unlearn Its Habits?
The attempt to “fix” this involves iterative prompting and the creation of secondary agents. By instructing an AI to act as its own editor—scanning for similes, cutting out vague words like “something” or “nowhere,” and enforcing strict rules on dialogue tags—the output improves.
One such internal set of instructions for mimicking Hemingway, for instance, involves stripping dialogue tags to bare “he said/she said” and utilizing short, declarative sentences connected by “and” to create forward momentum. While these constraints move the AI closer to a human-like output, they reveal the mechanical nature of the process: the AI isn’t “writing” so much as it is following a checklist of stylistic constraints.
Until LLMs can move beyond the statistical probability of the next word and understand the causal weight of a character’s choice, the “fidgeting” will remain. The machine can mimic the voice, but it still struggles to inhabit the room.
