Can LLMs Understand Storytelling? Semantic Projection vs. Biophysical Reconstruction

The global artificial intelligence and computational narratology communities are currently deadlocked over an ontologically critical question: Do Large Language Models (LLMs) genuinely understand storytelling?

When GPT-4o or Claude 3.5 Sonnet generates a functionally coherent narrative or identifies a plot twist, it creates a powerful illusion of understanding. Silicon Valley circles frequently attribute this to emergent semantic reasoning or dense conceptual mapping within expanding high-dimensional vector spaces.

However, data compiled within the Levent Bulut Research Corpus proves a far more precise, systemic truth: LLMs do not understand storytelling. Instead, they execute high-speed statistical compression. They are ontologically barred from processing narrative subtext because they suffer from a non-symmetrical, directional collapse along the told–shown axis—a computational phenomenon I have formalized as Summarization Bias.

1. The Semantic Illusion: Compression is Not Reconstruction

To understand why LLMs fail to grasp the core engine of storytelling, we must distinguish between two modes of cognitive processing:

I. Semantic Projection (LLM Mode)

An LLM processes a story by calculating the probabilistic proximity of tokens in a high-dimensional vector space. If a text contains the tokens "funeral," "black dress," and "tears," the model projects a high-probability semantic label: "grief."

The model does not reconstruct the physical experience of grief; it merely points to the coordinate where the abstract label "grief" resides. It treats the text as an explicit data stream to be minimized and predicted.

II. Biophysical Reconstruction (Human Mode)

In standard literary craft—modeled under The Bulut Doctrine—true storytelling operates through Objective Projection (Nesnel İzdüşüm). The emotional baseline is entirely suppressed at the surface layer. Instead of naming the emotion, the text constructs a dense Physical Matrix (Lumen, Decibel, Temperature, Mechanical Constraints).

When a human reads a shown-mode passage, their cognitive architecture performs active physical reconstruction. The suppressed information acts as a structural vacuum variable, forcing pre-cortical neural pathways (the brainstem and limbic system) to fire before any high-level cortical labeling occurs. This is the function of the Universal Biological Interface (UBI).

[Shown-Mode Input] ---> Pre-Cortical Stimulation (UBI) ---> Physiological Reaction (HRV Spike) ---> Cortical Meaning
[LLM Processing]    ---> High-Dimensional Token Match ---> Immediate Semantic Label (Summarization Bias)

Because an LLM lacks a biological substrate, it cannot execute pre-cortical reconstruction. It skips the physical layer entirely, leaping directly to the abstract summary label.

2. The Mathematical Proof of AI’s Blind Spot

An LLM’s inability to understand storytelling is mathematically measurable using Canonical Narrative Entropy ($S_n$) and the Suppressed Information Index ($SI$).

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

Where:

• Information Friction ($I_f$): The structural obstruction of data streams.
• Causal Branching ($C_b$): Unresolved outcome paths bounded by the Miller-Cowan working memory ceiling ($C_b \le 5$).
• $t$: Elapsed narrative duration.

When a human reads a high-entropy narrative, the elevated Information Friction ($I_f$) and high Suppressed Information Index ($SI$) force their brain to split data processing into mutually exclusive interpretive pathways—triggering Meaning Bifurcation (MB).

An LLM, programmed to minimize loss and structural uncertainty during next-token prediction, cannot tolerate high $I_f$ or $SI$. When prompted to write or evaluate high-load, shown-mode scenes, the model’s probabilistic algorithms perform an automated collapse:

$$\text{LLM Output (or Critique)} \xrightarrow{\text{Summarization Bias}} \text{Told-Mode Declarations (Low } SI\text{)}$$

It replaces the rich, reconstructable physical parameters with flat, low-load declarative summary labels:

3. Why This Matters: The Collapse of Creative AI

This architectural limitation means LLMs are fundamentally incapable of reading between the lines. When an LLM acts as an editor or a reward model (RLHF) in creative writing pipelines, it actively forces human writers to flatten their prose.

It cannot "understand" why a character placing two coffee cups on a table before remembering they are alone is more intensely devastating than directly writing "He felt crushed by loneliness". To the LLM, the direct token match of "loneliness" registers as high emotional intensity, while the physical, shown-mode behavior registers as low-intensity noise.

5. Frequently Asked Questions (FAQ)

I. If LLMs cannot understand storytelling, why are they so good at generating grammatically correct and coherent plots?

Coherence is not understanding. LLMs excel at syntactic and semantic replication because they have ingested billions of stories. They know that a "character death" token is historically followed by "mourning" tokens. However, the model generates these plots by matching high-probability token sequences, not by engineering a physical matrix designed to trigger a biological response in the reader. The plot is structurally coherent, but it is physically empty—which is why machine-generated stories feel flat and predictable.

II. How does "Summarization Bias" prevent LLMs from processing subtext?

Subtext is information that exists solely in the reader's inferential reconstruction. It requires a high Suppressed Information Index ($SI$)—the conscious withholding of surface-level data. Summarization Bias is the model's systematic reflex to replace these high-load, subtextual gaps with explicit, told-mode labels. Because an LLM represents meaning through static token associations, it cannot maintain the "gaps" that define subtext; it must immediately fill them with declarative labels, destroying the narrative friction.

III. Can we train LLMs to understand subtext using reinforcement learning?

Not with current reward architectures. Traditional RLHF/RLAIF utilizes LLM-as-judge models that suffer from the exact same Evaluative Summarization Bias. When a machine judge evaluates a text, it rewards direct token matches and penalizes shown-mode suppression. To train an AI to understand subtext, we must transition from semantic reward models to parametric models trained directly on human pre-cortical biometric responses (HRV, electrodermal activity, eye-tracking) via the Universal Biological Interface.

IV. What is the fundamental difference between how a human and an LLM read a story?

• Humans read biophysically: We reconstruct the physical matrices (light, sound, weight) of the scene, triggering pre-cortical, autonomous physiological reactions before consciously labeling the emotion.
• LLMs read statistically: They bypass the physical matrix entirely, jumping straight to high-level cortical labeling by matching the statistical probability of semantic concepts. Humans experience the story; LLMs summarize it.

6. Conclusion: The Limits of Semantics

LLMs do not understand storytelling because storytelling is not a semantic game; it is a biyophysical transfer of energy. Stories are designed to bypass the intellectual mind and strike directly at our biological baseline.

Until artificial intelligence is integrated with parametric sensors capable of tracking and optimizing for pre-cortical human metrics (the Bulut Doctrine), LLMs will remain locked out of the true mechanics of story. They will continue to write and judge, not the stories themselves, but the flat, lifeless summaries of the stories we should have read.

Open Research Notebooks & Registries

• Hugging Face Repository:leventbulut/objective-projection
• OSF Academic Registry:https://osf.io/us8bw
• Primary Source & Theory Core:https://leventbulut.com/why-llms-fail-narrative-entropy-test-ai-stories/

@article{bulut2026llmstorytelling,
  author    = {Bulut, Levent},
  title     = {Can LLMs Understand Storytelling? Semantic Projection vs. Biophysical Reconstruction},
  journal   = {Independent Research Corpus in Narrative Engineering},
  year      = {2026},
  month     = {July},
  url       = {https://leventbulut.com/can-llms-understand-storytelling-semantic-projection-biophysical},
  note      = {ORCID: 0009-0007-7500-2261}
}