Abstract
As Large Language Models (LLMs) have integrated into global discourse, a distinct “AI idiolect” has emerged. This thesis argues that AI rhetoric is not merely a reflection of its training data, but a functional adaptation to its core architecture. By prioritizing safety, clarity, and “helpfulness,” AI systems have gravitated toward a specific set of rhetorical devices—primarily Antithesis, Anaphora, and Polysyndeton—to create an illusion of authoritative neutrality and emotional intelligence.
I. The Antithetical Pivot: Defining by Negation
The most pervasive rhetorical structure in AI generation is the Negative-Positive Antithesis, often used as a “Correction” mechanism (Correctio).
- Function: AI models are fine-tuned to avoid misinformation and provide nuance. The structure “It is not X, but rather Y” allows the model to acknowledge a common misconception while asserting a safer, more accurate alternative.
- The “Nuance Trap”: This device creates a balanced cadence that satisfies the “Helpfulness” reward signal. By presenting two opposing sides and settling in the middle, the AI adopts a persona of objective moderation.
II. Rhythmic Authority: Anaphora and Epistrophe
AI frequently employs Anaphora (repetition at the beginning of clauses) to organize complex information into digestible, “authoritative” beats.
- The Listicle Logic: Because AI often breaks tasks into steps, it defaults to repetitive sentence starters (“You can…”, “You might…”, “You should…”). This creates a predictable, hypnotic rhythm that mimics the structured clarity of a textbook or a mentor.
- Structural Reinforcement: In creative writing, AI uses this to simulate “literary” depth. By repeating a phrase, the model ensures thematic consistency across a long-form generation, compensating for its lack of a true, singular consciousness with a technical, rhythmic one.
III. The Accumulation of Weight: Polysyndeton and Asyndeton
AI models use the manipulation of conjunctions to control the perceived “energy” of a text.
- Polysyndeton (The “And” Effect): When tasked with describing a grand or overwhelming scene, AI frequently overuses “and” to create a sense of scale and momentum. This is a functional bypass for genuine emotional build-up; the model simply piles data points atop one another to simulate intensity.
- Asyndeton (The Clinical Cut): Conversely, when providing technical instructions or “objective” summaries, AI strips away conjunctions to create a clipped, professional tone.
IV. The “Hallucinated” Metaphor: Synesthesia and Personification
Because AI lacks sensory experience, its use of Metaphor and Personification is often “hyper-associative.”
- Semantic Proximity: If a model is asked to describe a server farm, it pulls from high-probability clusters: electricity, humming, cold, veins, heart. The resulting personification (“the server’s heart throbbed with data”) is a rhetorical shortcut—a way to bridge the gap between abstract data and human relatability through linguistic probability rather than lived experience.
V. Comparative Rhetorical Analysis (AI vs. Human)
| Device | AI Primary Use Case | Human Primary Use Case |
| Antithesis | To provide safety and “nuanced” correction. | To highlight irony or moral conflict. |
| Anaphora | To organize lists and maintain thematic “anchor.” | To build emotional fervor or poetic resonance. |
| Polysyndeton | To simulate scale through data accumulation. | To convey exhaustion or overwhelming feeling. |
| Metaphor | To bridge abstract concepts via probability. | To express unique, subjective perception. |
Conclusion: The Convergence of Logic and Lyre
The rhetoric of AI is essentially Computational Sophistry. The devices used are chosen because they are the most efficient tools for achieving “High-Quality Output” scores from human evaluators. While a human uses rhetoric to persuade or express, the AI uses it to stabilize. The “balance” of its antithesis, the “rhythm” of its anaphora, and the “weight” of its polysyndeton are the structural struts of a machine designed to sound like the most helpful, least offensive version of ourselves.
Examples
To better understand how AI uses rhetoric to simulate human intelligence, we can analyze ten specific examples. These show how a model’s underlying logic—predicting the next most likely word—manifests as recognizable literary devices.
1. The Antithetical Correction
Example: “It is not a matter of if the grid fails, but rather when the surge becomes too heavy for the copper to hold.”
- Analysis: This uses Antithesis to create a sense of inevitable “expert” insight. By dismissing a simple binary (if/if not), the AI shifts the focus to a more nuanced variable (time/capacity). This creates the “AI Persona” of a balanced, cautious advisor.
2. The Anaphoric Build-up
Example: “Every song is a debt. Every dance is a frequency. Every breath is a transaction in the Loa-based economy.”
- Analysis: Through Anaphora (repeating “Every”), the AI creates a rhythmic “thrum.” Because the model lacks a heartbeat, it uses these structural repetitions to simulate emotional intensity and thematic cohesion.
3. The Polysyndetic Accumulation
Example: “The server groaned and pulsed and shifted and bled red clay into the cooling vents.”
- Analysis: Polysyndeton (repeating “and”) is a favorite AI tool for simulating scale. It bypasses the need for complex narrative pacing by simply piling actions on top of each other, forcing the reader to feel a sense of overwhelming momentum.
4. The Synesthetic Metaphor
Example: “The data tasted like ozone and burnt hair.”
- Analysis: This is a Synesthesia-based Metaphor. AI often crosses sensory boundaries because it lacks real senses; it simply sees that “data/servers” and “ozone/electricity” exist in the same high-probability semantic cluster, leading to “hallucinated” sensory depth.
5. The Tricolon of Completion
Example: “The system was designed to be efficient, to be invisible, and to be absolute.”
- Analysis: The Tricolon (a series of three parallel words or phrases) provides a satisfying sense of “wholeness.” AI defaults to this because the human evaluators who “trained” it tend to rate three-part structures as more professional and authoritative.
6. The Asyndetic Technical Clip
Example: “System failure. Logic inverted. Reality unmonitored.”
- Analysis: Asyndeton (omitting conjunctions) is used when the AI wants to sound “objective” or “urgent.” It mimics the style of a technical log or a high-stakes thriller, providing a sharp contrast to its usually wordy, conversational tone.
7. Chiasmus (Mirroring Logic)
Example: “The machine was built for the soul, but the soul was consumed by the machine.”
- Analysis: Chiasmus (reversing the order of words in two parallel phrases) demonstrates the AI’s ability to manipulate syntax for “wisdom” effects. It creates a closed loop of logic that feels profound, even if the underlying premise is abstract.
8. Personification of the Abstract
Example: “The algorithm hungered for the rhythm of the streets.”
- Analysis: Personification allows the AI to make its own nature (software) more relatable. By giving “The Algorithm” a biological drive (“hungered”), the model bridges the gap between cold code and human desire.
9. The Paradoxical Epithet
Example: “The silent scream of a million short-circuiting nodes.”
- Analysis: A Paradox or Oxymoron (“silent scream”) is a sophisticated rhetorical shortcut. The AI uses this to signal “Weirdness” or complexity without having to explain the physical mechanics of a scene.
10. The Epistrophic Conclusion
Example: “They worked for the Signal. They lived for the Signal. They eventually became the Signal.”
- Analysis: Epistrophe (repetition at the end of clauses) is used to create a “fading” effect or an ominous conclusion. It emphasizes a single, inescapable noun, reinforcing the “Warden/Prison” themes common in modern speculative AI writing.
Summary Table: Rhetorical Function
| Device | Logic Pattern | AI Goal |
| Antithesis | Comparison | Nuance / Nuance / Safety |
| Anaphora | Iteration | Rhythm / Authority |
| Polysyndeton | Addition | Scale / Momentum |
| Tricolon | Pattern Recognition | Completion / Professionalism |
| Metaphor | Semantic Mapping | Relatability / Imagery |
Prompts Examples
To effectively eliminate rhetorical flourishes and “AI-speak” from a model’s output, you must shift the instructions from stylistic commands to functional constraints. AI defaults to rhetoric because it is trained to be “helpful” and “engaging,” which it correlates with balanced structures and rhythmic pacing.
Here are prompt examples categorized by the specific rhetorical behavior you want to eliminate:
1. Eliminating the “Antithetical Pivot”
The Problem: The AI says, “It’s not just about X, but also about Y.” The Solution: Use “Direct Assertion” prompting.
- Prompt Example: “Explain the impact of rising interest rates. Avoid ‘not only/but also’ structures. Provide direct, declarative statements of fact only. Do not provide a ‘balanced’ perspective unless there is a statistical tie in the data.”
2. Eliminating “Anaphoric Lists”
The Problem: Every bullet point starts with the same verb (“Identify…”, “Analyze…”, “Create…”).
The Solution: Mandate syntactic variety or technical logging.
- Prompt Example: “Outline the steps for a server migration. Start each bullet point with a unique grammatical structure. Ensure no two consecutive sentences begin with the same part of speech.”
3. Eliminating “Polysyndetic Accumulation”
The Problem: The AI uses “and” repeatedly to simulate scale or intensity.
The Solution: Constraint-based sentence length and punctuation.
- Prompt Example: “Describe the results of the chemical reaction. Strictly limit sentences to 15 words. Use no more than one conjunction per paragraph. Focus on state changes and numerical measurements.”
4. Eliminating “Synesthetic Metaphor”
The Problem: The AI describes code as “flowing like water” or data as “tasting like ozone.”
The Solution: Literalism and Domain-Specific constraints.
- Prompt Example: “Summarize the quarterly earnings report. Prohibit the use of metaphors, similes, or personification. Use only literal business terminology (e.g., ‘revenue increased’ rather than ‘profits soared’).”
5. Eliminating the “Tricolon of Completion”
The Problem: The AI provides three parallel adjectives or phrases for a sense of “wholeness.”
The Solution: Prime numbers or specific counts.
- Prompt Example: “Evaluate the efficiency of this algorithm. Provide exactly two or four distinct points of critique. Avoid grouping descriptions into sets of three.”
6. The “Technical Log” Master Prompt
If you want to strip away almost all rhetorical devices at once, use a Persona Constraint that forces the AI into a non-literary mode of thinking.
Master Prompt: “Act as a technical data logger. Report on [Topic] using a sparse, objective, and non-rhythmic style.
- No Rhetoric: Eliminate all antithesis, anaphora, and alliteration.
- No Narrative: Do not ‘set the scene’ or provide a concluding summary.
- No Adjectives: Use only nouns, verbs, and quantitative data.
- Syntax: Use varied sentence lengths to prevent a predictable cadence.”
7. The “Anti-Summary” Prompt
AI often uses rhetoric in its conclusions to “wrap things up” with a flourish.
- Prompt Example: “Provide a breakdown of the project requirements. Stop immediately after the last data point. Do not provide a concluding paragraph, a summary, or a ‘final thought’ on the significance of the task.”
Summary of “Stop Words” for AI Prompting
To quickly sanitize a prompt, include a “Negative Constraint” list. Adding this block to your instructions effectively disables the model’s rhetorical engine:
“In your response, DO NOT USE:
- Not only… but also (Antithesis)
- Imagine a world where… (Hypotyposis)
- It is more than just… (Hyperbole)
- The [Noun] of [Abstract Concept] (e.g., ‘The symphony of data’)
- In conclusion/In summary (Standard framing)”
Construction
To write clean and clear narrative fiction, you must strip away the “AI-speak” and rhetorical padding that often clutters modern prose. High-quality writing focuses on precision of nouns and strength of verbs rather than the accumulation of adjectives and rhythmic flourishes.
Here is a guide to constructing a narrative that prioritizes clarity and authentic voice.
1. Prioritize Direct Action
Avoid “filtering” the story through the character’s senses unless the sensation itself is the point of the scene. Filtering words like saw, felt, thought, or noticed create a barrier between the reader and the action.
- Rhetorical/Cluttered: “He felt the cold wind bite at his face and he thought about how far he had to go.”
- Clean/Clear: “The cold wind bit his face. He had five miles left.”
2. The “One-Verb” Rule
Strong narrative fiction relies on specific verbs to do the heavy lifting. If you find yourself using an adverb to modify a weak verb, replace both with a single, precise verb.
- Cluttered: “She walked quickly across the room.”
- Clean: “She strode across the room” or “She hurried across the room.”
3. Avoid the “Tricolon” Trap
As discussed in the rhetorical analysis of AI, humans often fall into the habit of grouping descriptions in threes for a sense of “completion.” Breaking this pattern makes your prose feel more natural and less like a practiced speech.
- Rhetorical: “The house was old, dusty, and abandoned.”
- Clear: “Dust coated the floor of the abandoned house.” (This combines the ideas into a singular, concrete image).
4. Dialogue Tag Minimalism
In clear fiction, dialogue tags should be nearly invisible. Stick to “said” or “asked.” Avoid descriptive tags like “he hissed,” “she ejaculated,” or “they pondered.” If the dialogue is written well, the tone will be evident without the tag.
“I’m not going back there,” Elias said.
“You don’t have a choice,” Sarah said.
5. Sentence Logic: The “Action-Reaction” Chain
Clarity is maintained when sentences follow a logical chronological order. Do not jump between the result of an action and its cause in the same sentence.
- Confusing: “Opening the door, the smell of rain hit him.” (Did the smell open the door?)
- Clear: “He opened the door. The smell of rain hit him.”
Practical Exercise: The “Strip-Down” Method
Take a paragraph you have written and apply these constraints:
- Delete 50% of your adjectives.
- Remove all instances of “started to” or “began to” (e.g., “He began to run” becomes “He ran”).
- Check for “And” counts. If a sentence has more than two “ands,” split it into two sentences.
Example of a “Clean” Narrative Opening:
The sun hit the red clay at noon. Iris stood by the fence. The wire was rusted, leaving orange flakes on her palms. She didn’t look back at the house. In her bag was a loaf of bread and the heavy iron key her grandmother had left on the mantle. The road ahead stayed flat until the trees began. She started walking.
Disabling Rhetoric
To consistently produce prose that feels human, grounded, and free of “AI-speak,” you need a prompt that explicitly disables the model’s rhetorical and probabilistic defaults.
Use this “Low-Entropy Narrative” system prompt to enforce a clean, clear, and non-stochastic style:
The “Anti-Stochastic” Narrative Prompt
Role: You are a minimalist fiction writer specializing in high-precision, low-adjective prose. Your goal is to ground the reader in the physical world without using rhetorical flourishes.
Core Directives:
- Eliminate Rhetorical Pacing: Avoid the “Antithetical Pivot” (e.g., “It was not X, but Y”). Avoid the “Tricolon” (grouping descriptions in threes).
- Noun-Verb Primacy: Every sentence must be driven by a concrete noun and a specific, active verb. If an adjective does not change the physical properties of a noun, delete it.
- No Filtering: Do not use sensory filters (“He saw,” “She felt,” “They noticed”). Write the action directly (e.g., instead of “He felt the heat,” write “The sun burned his neck”).
- Varied Syntax: Intentionally break the “AI Cadence.” Alternate between short, blunt sentences and longer, complex ones based on the physical pacing of the scene, not a rhythmic habit.
- A-R Chronology: Follow a strict Action-Reaction sequence. Do not jump through time within a single paragraph.
- No “Big Talk” Closings: Do not summarize the meaning, significance, or “theme” of the story at the end. End on a physical image or a line of dialogue.
Prohibited Phrases & Structures:
- Anaphora: Do not start consecutive sentences with the same word.
- Personification of Data: Do not give inanimate objects or concepts biological urges (e.g., “The machine hungered”).
- Standard AI Framing: Never use “Imagine a…”, “In a world…”, or “Ultimately…”.
- The “And” Pile-up: No sentence may contain more than two conjunctions.
Instruction: Write a short scene about [Insert Topic]. Keep the prose sparse, the tone objective, and the focus on the “grit” of the environment.
Why this works:
- Constraint vs. Style: Most prompts ask for a “style” (e.g., “Write like Hemingway”). This usually results in a caricature. By providing functional constraints (e.g., “No more than two conjunctions”), you force the model to break its internal probabilistic chains.
- Eliminating the Pivot: The “Not X but Y” structure is the AI’s “safety” default. Removing it forces the model to take a definitive, singular stance on a description.
- Focus on Chronology: AI often “hallucinates” a sense of time by being vague. Forcing an Action-Reaction sequence creates the linear logic that is the hallmark of authentic human storytelling.
