How to Detect AI Writing: Methods, Tools & Technical Metrics - Geostar

Quick Detection Guide

Spot AI-generated text in seconds

Visual Scan

• ✗Excessive bullet points (groups of 3)
• ✗Perfect parallel structure
• ✗Em dashes everywhere (2+ per 200 words)
• ✗Uniform paragraph lengths
• ✗Bold headers + emojis

Phrase Check

Find 2+ of these = likely AI

• ✗ "delve into this topic"
• ✗ "tapestry of experiences"
• ✗ "navigate the landscape"
• ✗ "leverage synergies"
• ✗ "seamlessly integrate"
• ✗ "robust framework"
• ✗ "Here's why:" or "Here's what:"
• ✗ "It's not X. It's Y." (repeated)

Burstiness

1. Count words in 5 random sentences
2. Calculate: μ (average) and σ (std dev)
3. Formula: B = (σ / μ) × 100

B < 30 = Likely AI

Example: 24, 23, 26, 25, 24 → B = 4.7

B > 50 = Likely Human

Example: 8, 34, 12, 41, 15 → B = 66.6

20-Second Structure Test

Opening
"In today's rapidly evolving..." = AI

Body
Every paragraph same length? = AI

Lists
Groups of exactly 3 items? = AI

Closing
"What are your thoughts?" = AI

Rhythm
Metronome vs Jazz? Metronome = AI

Sentences
All 20-30 words? = AI

AI Detection Hierarchy

Confidence levels for identifying AI-generated text

Tier 1: Immediate Red Flags

90-99% AI probability

• Left-in prompt artifacts
• "Certainly, here is a possible introduction..."
• Left-in refusal text
• "I'm sorry, but I don't have access to real-time information"
• Multiple signature words
• "5+ words from signature list"
• 'delve' in academic context
• "We delve into the implications..."
• Perfect triadic repetition
• "5+ instances of groups of 3"

Tier 2: Strong Indicators

70-90% AI probability

• Low burstiness (B < 20)
• "All sentences 22-28 words"
• Uniform paragraph length
• "Every paragraph 4-5 sentences"
• Em dash frequency
• "5+ per 200 words"
• Setup phrase density
• "3+ 'Here's' or 'The truth is' per 300 words"
• Formulaic structure repetition
• "It's not X, it's Y' used 3+ times"

Tier 3: Moderate Indicators

40-60% AI probability

• Corporate buzzword density
• "3-5 words like 'leverage,' 'synergy'"
• Consistent Oxford commas
• "Never omitted"
• Question-answer cadence
• "2+ rhetorical Q&A pairs"
• Parallel bullet perfection
• "All bullets identical structure"
• Generic enthusiasm
• "Great question!" 'Excellent point!'

AI Detection Tools Comparison

Performance metrics for popular AI text detection tools

How to Defeat Detection

✓ What Works

• • Paraphrasing (-20% detection)
• • Humanizer Tools (90% effective)
• • Prompt Engineering (87% effective)
• • QuillBot (50% effective)

✗ What Fails

• • Removing Buzzwords (5% effective)
• • Minor Edits (10% effective)
• • Patterns remain structural, not lexical

Technical Detection Metrics

Burstiness Formula

Measures sentence length variation, the most reliable structural tell

Formula

B = (σ / μ) × 100

σ = Standard deviation

μ = Mean sentence length

B = Burstiness score

AI Pattern

B = 4.7

Sentences: 24, 23, 26, 25, 24 words

Average: 24.4

Std Dev: 1.14

Low variation = metronome rhythm = AI

Human Pattern

B = 66.6

Sentences: 8, 34, 12, 41, 15 words

Average: 22

Std Dev: 14.66

High variation = jazz rhythm = Human

Threshold: B < 30

Uniform sentence lengths indicate AI generation. Low burstiness reveals algorithmic consistency.

Threshold: B > 50

Natural variation in sentence length indicates human authorship. High burstiness reflects cognitive rhythm.

Perplexity Score Analysis

How "surprised" a language model is by word choices

Score 5-10

Extremely Predictable

AI-generated (GPT-4, Claude, Gemini)

Score 10-20

Very Predictable

Heavily edited or templated writing

Score 20-50

Normal Variation

Human writing

Score 50+

Unexpected/Creative

Literary fiction, poetry, technical jargon

Why AI scores low: Large language models generate the most probable next token. Predictability is the core mechanism. Detection tools like GPTZero use threshold >85 for likely human authorship.