Ai Writing Detector

Structural Tells

• Em-dash overuse: More than 1 per 200 words
• Rule of three: "X, Y, and Z" constructions in series
• Negative parallelisms: "not X, but Y", "not just...it's...", "not only...but also"
• Inline-header lists: Bullet points with bolded headers: followed by text

Content Patterns

• Legacy/symbolism puffery: "stands as a testament", "enduring legacy", "pivotal role"
• Superficial "-ing" analyses: Sentences ending with "...reflecting its importance" or "...highlighting the significance"
• Challenges sections: "Despite its [positive], [subject] faces challenges..."
• Hedging preambles: "It's important to note", "worth mentioning"

Style Markers

• Vague attribution: "Experts argue", "Observers note", "Industry reports suggest"
• Elegant variation: Avoiding word repetition by cycling synonyms unnaturally
• Promotional tone: "groundbreaking", "stunning", "nestled in the heart of"
• Hydrogen Jukebox: Abstract/concrete "eyeball kicks" crammed into every phrase (nostalgebraist, 2024)

Output Format

## AI Pattern Analysis: [filename]

### Summary
- **Confidence**: [Low/Medium/High] (based on pattern density)
- **Patterns found**: [count] across [categories] categories

### Vocabulary Markers ([count])
- Line X: "delve into the intricacies"
- Line Y: "pivotal role in fostering"

### Structural Patterns ([count])
- Em-dashes: [count] (rate: X per 1000 words)
- Rule of three: [count] instances
- Negative parallelisms: [count]

### Content Patterns ([count])
- Line X: Legacy puffery - "stands as a testament to..."
- Line Y: Superficial analysis - "...underscoring its significance"

### Flagged Passages
[Quoted passages with highest pattern density]

Scripts

Quick Check: check-ai-words

Vocabulary-only analysis with rate comparison.

check-ai-words <file> [multiplier]
# multiplier default: 3 (flag words at 3x base rate)

Full Analysis: check-ai-patterns

Comprehensive analysis including vocabulary, em-dashes, and phrase patterns.

check-ai-patterns <file> [multiplier]

Output includes:

• Em-dash rate per 1000 words (threshold: 5)
• Vocabulary violations with corpus base rates
• Common AI phrase detection
• Confidence score (0-7)

Data File: references/ai_word_rates.txt

Base rates per million words from ngrams.dev English corpus (23.6B words).

Detailed Pattern Reference

For comprehensive pattern lists with examples, see references/patterns.md.

Confidence Calibration

Single patterns are weak signals. Clusters of patterns, especially across categories, indicate AI generation.