AI Writing Patterns: Research & Data
We analyzed 186,000+ articles to map AI text patterns across 5+ models. Detection tools now identify AI writing with 97% accuracy using word choice and sentence rhythm alone. The data below covers everything content teams and SEO agencies need to know.
Rhetorical Formula Patterns by Model
Repetitive sentence structures that reveal AI writing across all major LLMs
Binary Contrast
“It's not X. It's Y.”
ChatGPT
Stop/Start Parallel
“Stop chasing leads. Start building relationships.”
ChatGPT
Question Negation
“The question isn't whether. It's whether your strategy reflects that.”
ChatGPT
Rule of Three
“fast, efficient, and user-friendly”
ChatGPT
Reframe Formula
“That's not failure. That's data.”
ChatGPT
Old/New Contrast
“Old approach: manual. New approach: automated.”
ChatGPT
False Comfort
“You may not control the market. But you can control your response.”
ChatGPT
Temporal Contrast
“We used to X. Now we Y.”
ChatGPT + Copilot
Concession + Pivot
“Yes, X. But also Y.”
Claude
Nuanced Hedge
“That's a fair point. The nuance is Y.”
Claude
Challenge Formula
“Popular opinion: X. Reality: Y.”
ChatGPT + Grok
Identity Redefinition
“I'm not a X. I'm a Y.”
ChatGPT (LinkedIn)
Parallel Contrast
“Done beats perfect.”
ChatGPT + LinkedIn
Escalation
“X is good. But Y is better.”
ChatGPT + DeepSeek
Drama Starters
“This changes everything. Nobody's talking about this.”
ChatGPT
Truth Announcement
“The truth is, customers don't care about features.”
All Models
Here's Setup
“Here's why this matters:”
All Models
Let That Sink In
“Let that sink in.”
ChatGPT (LinkedIn)
Setup Phrases & Transition Crimes
Instant AI tells: these phrases scream artificial writing
Setup Phrases That Scream AI
“Here's why:”
Fix: Start with the reason
“Here's what:”
Fix: Start with the content
“Here's the thing:”
Fix: Delete entirely
“What this means:”
Fix: State implication directly
“The reality is:”
Fix: State directly
“The truth is:”
Fix: Just be truthful
“Let's be honest:”
Fix: Delete, just be honest
“It's important to note that”
Fix: Just note it
“It's worth noting that”
Fix: State the fact
“Think about it like...”
Fix: State comparison directly
Transition Crimes
Moreover,
Better: Also, or new sentence
Furthermore,
Better: Additionally, or new sentence
Additionally,
Better: Start new thought
Consequently,
Better: So, Therefore, This means
That said,
Better: But or delete
That being said,
Better: Delete entirely
At the end of the day,
Better: Delete (meaningless)
The bottom line is,
Better: Delete, state it
On the flip side,
Better: But or However
With that in mind,
Better: Delete
AI Signature Words by Category
Overused vocabulary patterns organized by semantic function
Corporate Buzzwords
Empowerment Verbs
Aspirational Adjectives
Grandiosity Terms
Overused for minor changes
Rarely truly revolutionary
Usually has precedents
Overhyped claims
Syntactic & Punctuation Patterns
Syntactic Structure Tells
Sentence Length
AI Pattern
Uniform 25-27 words
Human Pattern
Highly varied (5-40+ words)
Detection: Calculate standard deviation
Burstiness
AI Pattern
Low (B < 30)
Human Pattern
High (B > 50)
Detection: B = (σ / μ) × 100
Present Participles
AI Pattern
2-5× human rate
Human Pattern
Occasional
Detection: Count -ing verbs
Nominalizations
AI Pattern
1.5-2× frequency
Human Pattern
Moderate use
Detection: 'implementation' vs 'implement'
Passive Voice
AI Pattern
GPT-4o: half human rate
Human Pattern
Varies by context
Detection: Count 'was done' vs 'did'
Coordination
AI Pattern
Higher frequency
Human Pattern
Balanced with subordination
Detection: Multiple 'and' connections
Punctuation Usage Patterns
Em dash (—)
AI Behavior
10× increase GPT-3.5→GPT-4o
Human Behavior
Occasional, varied
Fix
Replace with comma, period, or colon
Oxford comma
AI Behavior
Almost never omitted (100%)
Human Behavior
Inconsistent usage
Fix
Vary (but maintain style guide)
Colon-reveal
AI Behavior
Result: 80% fewer
Human Behavior
Rare in body text
Fix
Two sentences
Exclamation marks
AI Behavior
Conservative, formula-based
Human Behavior
Emotional, spontaneous
Fix
Use sparingly, naturally
Purple Prose: AI Creative Writing Patterns
Why AI defaults to ornate, melodramatic language in fiction and creative content
Comparative Examples
AI-Generated
“The wizard approached cautiously, his robes flowing like purest velvet, the shimmering fabric catching ethereal moonlight as it cascaded down in waves of mystical elegance.”
Problem: Excessive elaboration with stacked adjectives and overwrought metaphors
Human Revision
“The wizard's velvet robes shimmered in the moonlight.”
Fix: Cut adjective clusters, remove redundant metaphors, focus on concrete details
AI-Generated
“Her heart was a tapestry of emotions, woven with threads of joy and sorrow, creating a symphony of feelings that resonated through her very soul.”
Problem: Multiple generic metaphors (tapestry, symphony, soul) instead of specific emotion
Human Revision
“She felt both happy and sad.”
Fix: Name the actual emotions, skip the metaphorical language
AI-Generated
“Tears streamed down her face like diamonds, each one a precious testament to the depth of her overwhelming anguish.”
Problem: Melodramatic simile and emotional amplification
Human Revision
“Tears streamed down her face.”
Fix: Simple, direct description without comparison
AI-Generated
“The changeling's fur was soft as butterfly wings, shimmering with an otherworldly luminescence that seemed to pulse with ancient magic.”
Problem: Stacked similes and unnecessary elaboration
Human Revision
“The changeling's soft fur glowed.”
Fix: One concrete detail, delete the rest
Why AI Defaults to Purple Prose
RLHF Evaluator Preference
Human evaluators in training preferred 'literary' language, rewarding elaborate vocabulary over plain prose.
Training Data Bias
Overrepresentation of published fiction (which is already more ornate) vs. rough drafts or conversational writing.
Register Leveling
Models apply formal, elaborate style universally — fiction, emails, and technical writing all get the same treatment.
Metaphor Overuse
Generic metaphors like 'tapestry' and 'symphony' appear in ~8% of creative outputs due to high training frequency.
Generic AI Character Names
Frequency analysis of character names in AI-generated fiction reveals predictable patterns
Emily
60-70% frequency — most common in training data
Sarah
60-70% frequency — most common in training data
James
High frequency — generic Anglo name
Michael
High frequency — generic Anglo name
Emma
Common contemporary name in training corpus
David
Generic biblical/traditional name
How to Spot Purple Prose
Look for: Elaborate adjectives before every noun ("the ethereal moonlight," "the tumultuous emotions"), metaphor clusters ("a tapestry of memories woven through the fabric of time"), and emotional amplification ("her heart shattered into a million pieces").
Why it happens: RLHF training on creative writing rewards "literary" language. Evaluators prefer ornate prose over plain language, creating a feedback loop that amplifies melodrama and complexity.
Fix: Replace adjective clusters with single concrete details. Cut metaphors by 80%. Use one-syllable verbs. Delete emotional amplifiers.
Want to get cited in AI search?
Geostar is a GEO and SEO agency. Our team handles content, technical implementation, and web development to get your brand cited in ChatGPT, Perplexity, and Google AI Overviews.
Brands that act now will own the next era of search.