Claude-Specific Vocabulary Database

Claude-Specific Vocabulary Database

A comprehensive database of vocabulary, phrases, and patterns that are statistically overrepresented in Claude outputs. Use this to detect and humanize AI-generated text.

How Detection Works

AI detection systems measure several statistical properties:

1. Lexical Fingerprints: Every LLM has vocabulary preferences—words it uses 2-10x more frequently than humans
2. Perplexity Patterns: AI text has unnaturally consistent complexity; humans vary wildly
3. Burstiness: Humans write in bursts (short-short-LONG-short); AI tends toward regularity
4. Semantic Coherence: AI text is often too coherent—real posts have tangents

This database addresses #1: lexical fingerprints.

CATEGORY 1: Formal Transition Words

Claude over-indexes on formal connectors. Replace with casual alternatives or delete entirely.

CATEGORY 2: Filler Hedges (Overly Polite)

Claude hedges excessively. Humans are more direct.

CATEGORY 3: Formal Verb Choices

Claude uses formal verbs. Humans use simpler ones.

CATEGORY 4: Formal Noun Choices

CATEGORY 5: Prepositional Phrases (Over-Formal)

CATEGORY 6: Claude Power Words (Overused)

These words appear 3-10x more frequently in Claude outputs than in human Reddit posts.

CATEGORY 7: Claude Phrase Patterns

Full phrases that are Claude signatures.

"Happy To" Patterns (Delete All)

• "I'd be happy to help with..."
• "I'm happy to explain..."
• "I'd be glad to assist..."
• "I'm pleased to..."

Fix: Just do the thing. Never announce willingness.

"Let Me" Patterns (Delete All)

• "Let me explain..."
• "Let me share..."
• "Let me walk you through..."
• "Let me break this down..."

Fix: Just explain/share/break it down. No announcements.

"It's Important To" Patterns

• "It's important to note that..."
• "It's worth mentioning that..."
• "It's crucial to understand that..."
• "It's essential to recognize that..."

Fix: Just state the thing. Or: "honestly" / "btw" / "oh and"

"This Is" Generalizations

• "This is a common problem..."
• "This is understandable..."
• "This is a great question..."

Fix: Delete or be specific: "I had this same thing" / "yeah" / [just answer]

Journey/Process Language

• "on this journey"
• "throughout this process"
• "along the way"
• "as you move forward"

Fix: Delete entirely. Real people don't use "journey" for mundane things.

CATEGORY 8: Opening Patterns (Instant Flags)

These openings immediately signal AI:

Better Openings (Human Style)

• Just start with the topic
• "ok so basically..."
• "honestly..."
• "ngl..."
• "so here's the thing..."
• "yeah so..."
• Story/anecdote first

CATEGORY 9: Closing Patterns (Instant Flags)

Better Closings (Human Style)

• Trail off with uncertainty: "idk if that helps..."
• Self-doubt: "but yeah that's just my experience"
• Casual: "anyway"
• Just stop writing
• "hope this makes sense lol"

CATEGORY 10: Structural Patterns

Sentence Length

AI Pattern: Consistent 15-25 words per sentence Human Pattern: Wild variance (3 words to 40 words)

Test: Calculate standard deviation of sentence lengths. Should be > 8 words.

Paragraph Structure

AI Pattern: 3-4 sentences per paragraph, consistently Human Pattern: 1 sentence, then 5 sentences, then 2 sentences

Test: No two consecutive paragraphs should be the same length.

List Usage

AI Pattern: Numbered lists, parallel structure, 3-5 items Human Pattern: Bullet points are rare; when used, inconsistent formatting

Test: If you have a list, break it. Vary bullet styles. Don't align.

Transitions Between Paragraphs

AI Pattern: Every paragraph logically connects to the previous Human Pattern: Topic jumps, "oh and I forgot to mention," tangents

Test: At least one paragraph should feel like a non-sequitur.

CATEGORY 11: Emotional Language Flags

CATEGORY 12: Reddit-Specific Humanizers

Add these naturally to pass as human:

Casual Fillers (Need 3+ per post)

• honestly
• tbh
• like
• idk
• ngl
• or something
• I guess
• maybe
• kinda
• sorta
• basically
• pretty much
• I mean

Reddit Speech Patterns

• "Edit: ..." (to add something)
• "lol" / "lmao" (for mild humor)
• "idk if this helps but..."
• "might be wrong but..."
• "not sure if this is the right sub but..."

Uncertainty Markers

• "I think?"
• "maybe?"
• Trailing "..."
• "or maybe not idk"
• "but who knows"

Self-Deprecation

• "I'm probably doing this wrong but..."
• "feel free to tell me I'm an idiot"
• "this is probably obvious but..."

USAGE INSTRUCTIONS

During Detection Scan (Phase 4)

1. Run vocabulary check: Search for every item in Categories 1-6
2. Flag and replace: Use human alternatives
3. Check phrase patterns: Search for Category 7-9 patterns
4. Delete instead of replacing when in doubt
5. Add humanizers from Category 12 if count < 3

Statistical Targets

EXPANDING THIS DATABASE

When you notice new Claude-isms:

1. Check frequency in AI vs human text
2. Find 2-3 human alternatives
3. Add to appropriate category
4. Test replacement in context

This is a living document.