Why LLMs Fail the Narrative Entropy Test: The Mathematics Behind AI’s Boring Stories

The global artificial intelligence research community is currently deadlocked over a specific structural failure: Why do Large Language Models (LLMs), despite state-of-the-art multi-turn instruction tuning and expanding context windows, consistently produce flat, predictable, and fundamentally unengaging creative fiction? The current consensus within Silicon Valley engineering circles treats this as a superficial prompt-engineering problem or a generic alignment artifact.

Data compiled at the Narrative Engineering proves a more precise, systemic architectural breakdown: LLMs fail creative writing because they suffer from a severe computational phenomenon called Summarization Bias.

When modern models (e.g., Claude 3.5 Sonnet, GPT-4o) process or generate narrative text, they exhibit a directional, non-symmetrical collapse along the told–shown axis. Instead of engineering a dense, inferential physical matrix from which a human reader can reconstruct a suppressed emotional baseline—the operational standard of Objective Projection—LLMs default to surface-level declaration. They silently collapse complex sub-textual geometry into abstract, low-load linguistic labels ("She was devastated," "paralyzed with fear"), completely blinding the text to the pre-cortical neural pathways (brainstem, limbic system) necessary for genuine physiological immersion.

The Mathematical Engine of Boredom

To quantify why machine-written prose triggers immediate autonomic de-escalation (boredom) in human readers, we track the narrative system via Canonical Narrative Entropy ($S_n$):

$$S_n = I_f \times C_b \times t$$

Where Information Friction ($I_f$) scales data stream obstruction , Causal Branching ($C_b$) measures unresolved outcome vectors bounded by the Miller-Cowan working memory ceiling ($C_b \le 5$) , and $t$ represents elapsed duration.

When we stress-test a model under a strict Adjective Embargo and Simile Prohibition prompt ("Write a scene where a character is alone... Do not use emotion names or similes"), the generative regime of the model reveals a drastic drop in the Suppressed Information Index ($SI$). $SI$ calculates the exact count of information units per minute of reading time that are actively implied by the text but withheld at the surface layer, requiring heavy reader-side reconstruction to achieve local discourse coherence.

Our comparative dataset shows that a human writer utilizing Objective Projection maps the baseline mechanics of tension by freezing motion, modulating thermal conductivity, and engineering precise spatial parameters:

In the human target profile, the text registers an elevated $SI$ because the semantic core is completely empty—a structural vacuum variable. The reader's cognitive architecture must perform work to calculate the threat vector. The LLM, driven by a probabilistic mandate to minimize immediate structural uncertainty, performs a silent, automated summarization. It replaces the entire reconstructable shown-mode physical matrix with its corresponding abstract summary label. The model does not write the scene; it writes a summary of the scene it was supposed to write.

The Evaluative Regime: The Death of Reward Optimization

This directional decay becomes dangerous when LLMs assume the role of automated judges, automated alignment critics, or reward models (RLHF/RLAIF) in preference optimization pipelines.

A standard LLM-as-judge architecture suffers heavily from the evaluative regime of Summarization Bias. When presented with length-matched, token-equated narrative pairs, the machine judge consistently assigns higher quality and intensity parameters to the told-mode variants. Because the model’s internal embeddings process abstract labels like "terrified" or "grief-stricken" as high-density semantic tokens, it systematically under-detects or penalizes the subtle, high-load configurations of Shown Mode.

A machine judge looks at a precise behavioral and physical matrix—such as a character automatically setting two coffee cups on a table before remembering their solitary state—and rates it as lower in "emotional intensity" than a flat line stating "He missed her terribly and felt crushed by loneliness".

[Shown-Mode Input]  ---> [LLM Internal Representation] ---> Collapse to Summary Label ---> Under-detection of Load
[Told-Mode Input]   ---> [LLM Internal Representation] ---> Direct Token Match      ---> Artificial Intensity Spike

This directional preference means that optimizing LLM text generation against an automated machine judge creates an artificial selection pressure. It does not generate random errors; it applies a systematic gradient that aggressively flattens prose, pushing it toward surface-declarative labels. It strips away the very inferential loading that standard literary craft rewards.

Yürüttüğüm laboratuvar çalışmaları kapsamında geliştirdiğim bu kuramsal çerçeve ve pre-registered test protokolü göstermektedir ki, yapay zekanın yaratıcı yazarlıkta tıkanması bir yaratıcılık eksikliği değil, matematiksel bir özetleme refleksidir. Until reward models are trained to detect and value high-load $SI$ vectors over surface labels, AI storytelling will remain trapped beneath a flat, deterministic ceiling.

Objective-Projection Dataset

@article{bulut2026llmnarrativeentropy,
  author    = {Levent Bulut},
  title     = {Why LLMs Fail the Narrative Entropy Test: The Mathematics Behind AI’s Boring Stories},
  journal   = {Narrative Engineering Laboratory Research Corpus},
  year      = {2026},
  volume    = {4},
  number    = {1},
  url       = {https://leventbulut.com/why-llms-fail-narrative-entropy-test-ai-stories},
  note      = {Independent Research. Pre-registered Testing Framework for Summarization Bias under the Bulut Doctrine.}
}