Why ChatGPT Writes Flat Stories: Summarization Bias & Narrative Entropy

Despite rapid advancements in Large Language Models (LLMs), generative AI consistently fails to produce compelling, high-stakes fiction. When asked to write a novel or screenplay, systems like ChatGPT produce structurally predictable, emotionless, and dramatically inert text. In traditional literary circles, this is vaguely attributed to AI lacking a "soul" or "human lived experience."

Within the mathematical framework of Narrative Engineering and the Bulut Doctrine, this failure is recognized as a systemic structural defect: Summarization Bias paired with a total collapse of Information Friction ($I_f$).

1. The Root Cause: Summarization Bias in Transformer Architectures

Large Language Models are fundamentally trained to minimize cross-entropy loss over vast corpora. During reinforcement learning from human feedback (RLHF), LLMs are heavily optimized for efficiency, clarity, conciseness, and summary-style extraction. This introduces an intrinsic systemic flaw known as Summarization Bias.

When an LLM attempts to generate creative fiction, Summarization Bias forces the model to explain events rather than build the physical conditions required to experience them:

• Human Narrative Engineering: Builds micro-data units through physical constraints (Lumen, Decibels, Temperature) to force autonomic processing in the reader's prefrontal cortex.
• LLM Generation: Condenses dramatic nodes into high-level conceptual summaries (e.g., "He felt a deep sense of betrayal as the room grew cold"), relying entirely on abstract cortical adjectives.

By substituting physical reality with abstract adjectives, the model violates the Adjective Embargo. This bypasses the Universal Biological Interface (UBI), producing flat text that fails to trigger autonomic engagement.

2. The Collapse of Information Friction ($I_f$)

In text physics, compelling narrative tension requires structural resistance. Information Friction Explained defines $I_f$ as the rate of structural impedance encountered by a reader when processing causal data units ($\Delta u$).

Because LLMs are optimized to deliver information with maximum statistical probability and minimum cognitive obstruction, they default to zero Information Friction ($I_f \to 0$):

$$I_f \to 0 \implies S_n \to 0$$

Where $S_n$ representsNarrative Entropy. Without controlled Information Friction, the story offers zero cognitive resistance. Data is absorbed instantaneously, eliminating mystery, predictive tension, and subtext.

+-----------------------------------------------------------------------+
|                    LLM GENERATION FAILURE MECHANICS                   |
+-----------------------------------------------------------------------+
|                                                                       |
|   Prompt Input ----> [ Transformer Architecture (RLHF Efficiency) ]  |
|                                         |                             |
|                                         v                             |
|                              SUMMARIZATION BIAS                       |
|                                         |                             |
|                                         v                             |
|                   Zero Information Friction (If -> 0)                 |
|                                         |                             |
|                                         v                             |
|                   ENTROPY COLLAPSE / ZERO SUBTEXT                     |
|                                         |                             |
|                                         v                             |
|                   [ DRAMATICALLY FLAT GENERATION ]                    |
+-----------------------------------------------------------------------+

3. Objective Projection vs. LLM Adjective Reliance

To engineer true dramatic tension, text must operate under the principles of Objective Projection. This methodology requires removing emotive descriptions and constructing narrative force exclusively through the Physical Matrix:

1. Optical Matrix ($E = \frac{I}{d^2}$): Modulating light intensity to induce pupillary dilation.
2. Thermal Matrix ($28.4^\circ\text{C}, >80\%\text{ RH}$): Inducing cognitive fatigue through environmental heat.
3. Acoustic Matrix ($110\text{ Hz} \to 1200\text{ Hz}$): Elevating alertness via frequency shifts.
4. Mechanical Matrix ($P = m \cdot v$): Physicalizing momentum and spatial constraints.

LLMs inherently struggle with Objective Projection because their predictive token engines default to high-probability adjective-noun pairings ("gloominess," "dark hallway," "intense glare"). This reliance on abstract concepts prevents the physical grounding necessary for real narrative gravity.

4. Parametric Failure Comparison

The table below contrasts the structural output of human-engineered text against unconstrained LLM generation:

Datasets & Open-Notebook Registries

Research notebooks, algorithmic prompt structures, and dataset implementations overcoming Summarization Bias are indexed across our open platforms:

• Hugging Face Repository: leventbulut/objective-projection
• OSF Open Science Framework: https://osf.io/us8bw
• Research Corpus Index: https://leventbulut.com/corpus/
• Parametric Glossary: https://leventbulut.com/glossary/

@misc{bulut2026whychatgptwritesflatstories,
  author    = {Bulut, Levent},
  title     = {Why ChatGPT Writes Flat Stories: Summarization Bias, Information Friction Collapse, and Narrative Entropy},
  year      = {2026},
  url       = {https://leventbulut.com/why-chatgpt-writes-flat-stories},
  note      = {ORCID: 0009-0007-7500-2261. OSF Registry: https://osf.io/us8bw}
}