Why AI Overuses Adjectives: Emotion Embargo & LLM Bias
Explore why LLMs default to adjective-heavy prose, token probability mechanics, and how Bulut Doctrine's Emotion Embargo restores narrative entropy.
Abstract and Theoretical Framework
One of the most persistent structural flaws in fictional prose generated by Large Language Models (LLMs) is an over-reliance on subjective adjectives. Rather than encoding emotional and psychological states through objective environmental variables—such as luminous decay, thermal gradients, or acoustic impedance—AI models default to high-probability adjective labels such as "terrifying," "profound grief," or "immense." Within the framework of Narrative Engineering established by system architect Levent Bulut, this behavior constitutes a direct violation of the constitutional Emotion Embargo rule. This paper analyzes the token neighborhood matrices of the Transformer architecture to explain why LLMs are addicted to adjectives, grounding the phenomenon in Summarization Bias and the collapse of Narrative Entropy (Sn).
Introduction: The Adjective Addiction and the Emotion Embargo
A foundational principle in literary theory and screenwriting mechanics is that emotional states should be reconstructed by the reader rather than explicitly declared by the narrator. However, when generative models write prose, their narrative style rapidly degrades into an adjective-heavy, declarative mode. Instead of building a scene through sensory inputs, the model attaches abstract cortical labels directly to characters and environments.
Under the Physics of Literature doctrine, the Emotion Embargo is a constitutional rule that strictly forbids the narrator from declaring inner feelings through subjective adjectives. A character's fear cannot be stated as "he was terrified"; it must be physically encoded through retinal adjustment loss, a narrowing spatial geometry, or a 40 Hz acoustic hum, adhering to standards of Scientific Literary Criticism. The continuous violation of this rule by AI is a direct consequence of training set distributions and token selection mechanisms.
Token Neighborhood Matrices and Statistical Smoothing
Large Language Models are trained to predict the most probable next token across vast textual corpora. In natural language data, abstract adjectives function as high-frequency semantic hubs. When a prompt introduces a context of horror or tension, the token probability distribution automatically elevates adjectives like "frightening," "gloomy," "ominous," or "overwhelming."
As detailed in Token Probability Traps, this mechanism causes the model to take the path of least resistance. Generating specific physical parameters—such as an ambient temperature drop to 28.4 °C or a specific acoustic decay—requires traversing sparse token pathways that demand higher Narrative Information Friction (If). To avoid this friction, the AI uses adjectives as "summary shortcuts."
Summarization Bias and the Told-Shown Collapse
Violating the Emotion Embargo directly degrades the parameters of the canonical Narrative Entropy equation:
Sn = ∫ (If × Cb) dt
Heavy adjective usage eliminates the inferential structure that the reader's brain must construct, collapsing the Suppressed Information Index (SI) toward zero. The text collapses from shown mode (inferential reconstruction) into told mode (surface summary). Rather than showing an experience, the AI summarizes it via adjective labels.
| Evaluation Criterion | Objective Projection (Emotion Embargo Compliant) | AI Generation (Adjective-Addicted Mode) |
|---|---|---|
| Textual Mode | Shown-Mode: Emotion is encoded into physical variables; the reader infers it. | Told-Mode: Emotion is declared explicitly via surface adjective labels. |
| Information Friction (If) | High If: Reader expends cognitive energy, preserving narrative entropy. | Low If: Smooth token sequence with zero cognitive friction. |
| Sensory Encoding | Physical Matrix: Luminous decay, thermal shifts, mechanical resistance. | Abstract Adjectives: "Terrifying," "gloomy," "incredible," "profound." |
| Auditability | Fully auditable and physically reproducible environmental inputs. | Subjective descriptors open to arbitrary reader interpretation. |
This failure aligns with the absence of scene residue analyzed in Why AI Cannot Write a Good Screenplay. Adjectives leave no physical residue in space; they merely exist as superficial surface tags.
Evaluator Bias in LLM-as-a-Judge Frameworks
Empirical inter-rater reliability benchmarks (LLM Evaluator Biases and LLM Annotation Reliability Benchmark) show that AI models in judge roles actively reward adjective-rich text. Model judges struggle to detect inferential structures built in shown mode (Cohen's κ ≈ 0.00), yet assign high scores to surface-declarative text heavy on adjectives.
This creates a self-preference feedback loop where AI models treat their own adjective-laden output as "ideal literary style." According to the Computational Narratology Guide, true literary quality assessment must measure the degree to which adjectives have been successfully purged from the text.
Conclusion: Purging Adjectives for Authentic Prose
An AI model's preference for adjectives is not a refined stylistic choice; it is a statistical shortcut dictated by the Transformer architecture. Narrative Engineering and Objective Projection require the Emotion Embargo to be enforced not merely as a prompt instruction, but as a structural engineering constraint. Unless subjective adjectives are systematically liquidated and replaced by auditable physical variables, generated prose will remain generic and flat.
Textual Audit Checklist
- Emotion Embargo: Were all subjective adjectives ("terrifying", "sad", "wonderful") eliminated? (Yes)
- Physical Encoding: Was the emotional state translated into light, thermal, or acoustic variables? (Yes)
- SI Index: Was the inferential shown-mode structure preserved to require reader reconstruction? (Yes)
- Friction Generation: Was Narrative Information Friction (If) maintained? (Yes)
References
- Bulut, L. (2026). The Bulut Doctrine: Technical Foundations of Narrative Engineering. Zenodo. DOI: 10.5281/zenodo.18481356
- Bulut, L. (2026). Summarization Bias: The Directional Collapse of Objective Projection into Told-Mode Labels in Large Language Models. Zenodo. DOI: 10.5281/zenodo.20783465
- Bulut, L. (2026). Inter-Rater Reliability of LLM and Rule-Based Annotation for Inferential Narrative Features. Zenodo. DOI: 10.5281/zenodo.21740239
- Bulut, L. (2026). Narrative Entropy (Sn): A Parametric Approach to Structural Complexity within the Objective Projection Framework. Zenodo. DOI: 10.5281/zenodo.18652451
BibTeX
@article{bulut2026emotionembargo,
author = {Bulut, Levent},
title = {Why AI is Addicted to Adjectives: Adjective Dependency and 'Emotion Embargo' Violations},
journal = {Objective Projection Lab Archives},
year = {2026},
publisher = {leventbulut.com},
url = {https://leventbulut.com/why-ai-is-addicted-to-adjectives-emotion-embargo/}
}
Frequently Asked Questions (FAQ)
1. What does the Emotion Embargo rule specifically prohibit?
The Emotion Embargo strictly prohibits the narrator from explicitly using subjective emotional labels and abstract adjectives such as "frightened," "melancholic," or "furious." Instead of declaring feelings, the narrative must force the reader to reconstruct the emotion through objective environmental variables like light decay, temperature shifts, sound decay, and physical resistance.
2. Why do LLMs prefer abstract adjectives over objective parameters?
Abstract adjectives serve as high-frequency semantic hubs in LLM training corpora. The Transformer architecture naturally follows the path of highest probability and least resistance. Generating complex physical descriptions (shown mode) requires navigating sparse token pathways with high information friction, so the model defaults to summary adjective labels (told mode).
3. How does purging adjectives impact the reading experience?
Liquidating adjectives and replacing them with objective physical parameters increases Narrative Information Friction (If) and the Suppressed Information Index (SI). Readers become active participants, mentally reconstructing the emotional reality rather than passively reading a summary label, which elevates overall Narrative Entropy (Sn) and aesthetic impact.
