ai-writing-detection（AI 书写检测）

AI Writing Detection Reference

Expert-level knowledge base for detecting AI-generated text, compiled from academic research, commercial detection tools, and empirical analysis.

Quick Reference: High-Confidence Signals

These indicators strongly suggest AI authorship when found together:

Vocabulary Red Flags

High-signal words (50-700x more common in AI text):

• "delve", "tapestry", "nuanced", "multifaceted", "underscore"
• "intricate interplay", "played a crucial role", "complex and multifaceted"
• "paramount", "pivotal", "meticulous", "holistic", "robust"
• "stands/serves as", "marking a pivotal moment", "underscores its importance"

Overused phrases:

• "It's important to note that..."
• "In today's fast-paced world..."
• "At its core..."
• "Without further ado..."
• "Let me explain..."

See reference/vocabulary-patterns.md for complete lists.

Structural Red Flags

• Uniform sentence lengths: 12-18 words consistently (low burstiness)
• Tricolon structures: "research, collaboration, and problem-solving"
• Em dash overuse: AI uses em dashes in a formulaic way to mimic "punched up" sales writing, especially in parallelisms ("it's not X — it's Y"); swapping punctuation doesn't fix the underlying emphasis pattern
• Perfect paragraph uniformity: All paragraphs same approximate length
• Template conclusions: "In summary...", "In conclusion..."
• Negative parallelisms: "It's not about X; it's about Y"
• Elegant variation: Cycling through synonyms to avoid repetition
• False ranges: "From X to Y" with incoherent endpoints

See reference/structural-patterns.md for details.

Content Red Flags

• Importance puffery: "marking a pivotal moment in history"
• Ecosystem/conservation claims without citations
• "Challenges and Future" sections following rigid formula
• Promotional language: "nestled in", "stunning natural beauty", "boasts"
• Superficial analyses: "-ing" phrases attributing significance to facts

See reference/content-patterns.md for details.

Formatting Red Flags

• Title Case in all section headings
• Excessive boldface (every key term bolded)
• Inline-header lists: **Bold Header**: description pattern
• Emojis in formal content or headings
• Subject lines in non-email contexts

See reference/formatting-patterns.md for details.

Markup Red Flags (Definitive)

• turn0search0, turn0image0: ChatGPT reference markers
• contentReference[oaicite:]: ChatGPT reference bugs
• utm_source=chatgpt.com: URL tracking (definitive)
• Markdown in wikitext: ## headers, bold, text
• grok_card XML tags: Grok/X specific

See reference/markup-artifacts.md for details.

Citation Red Flags

• Broken external links that never existed (no archive)
• Invalid DOIs/ISBNs: Checksum failures
• Declared but unused references: Cite errors
• Placeholder values: url=URL, date=2025-XX-XX

See reference/citation-patterns.md for details.

Tone Red Flags

• Passive and detached voice throughout
• Absence of first-person pronouns where expected
• Consistent formality with no stylistic variation
• Over-politeness and excessive hedging

Detection Methodology

Multi-Layer Analysis Approach

Layer 1: Technical Artifact Scan (Definitive)

• Check for turn0search/oaicite markers (ChatGPT)
• Check for utm_source=chatgpt.com in URLs
• Check for grok_card tags (Grok)
• Check for Markdown in non-Markdown contexts
• If found: Definitive AI involvement

Layer 2: Vocabulary Pattern Matching

• Scan for overused AI words/phrases
• Count frequency of flagged terms
• Look for clusters of high-signal vocabulary
• Check for importance/symbolism phrases

Layer 3: Structural Analysis

• Observe sentence length variation (uniform = AI signal)
• Check paragraph uniformity
• Identify repetitive syntactic templates (tricolons, negative parallelisms)
• Look for elegant variation (synonym cycling)
• Check for false ranges

Layer 4: Content Pattern Analysis

• Check for importance puffery and promotional language
• Look for "Challenges and Future" formula
• Check for ecosystem/conservation claims without citations
• Identify superficial analyses with "-ing" attributions

Layer 5: Citation Verification

• Test external links - do they exist?
• Verify DOI/ISBN checksums
• Check for declared but unused references
• Look for placeholder values

Layer 6: Formatting Analysis

• Check heading capitalization (Title Case = signal)
• Count bold phrases per paragraph
• Look for inline-header list patterns
• Check for emojis in formal content

Layer 7: Stylometric Observation

• Pronoun usage patterns (missing first-person?)
• Tone consistency (too uniform = AI signal)
• Punctuation patterns (em dash overuse? curly quotes?)

Layer 8: Coherence Check

• Do paragraphs build a coherent argument?
• Are concepts repeated with different words?
• Do transitions actually connect ideas?

Layer 9: Confidence Scoring

• Weight multiple signals together
• Require corroborating evidence (3+ signals minimum)
• Apply context-specific adjustments
• Check for mitigating factors (human signals)
• Consider ineffective indicators (don't use them)

Model-Specific Patterns

Different AI models have distinct "fingerprints":

Important dates:

• ChatGPT launched: November 30, 2022 (text before this is almost certainly human)
• "delve" usage dropped: 2025 (still signals pre-2025 ChatGPT)

See reference/model-fingerprints.md for detailed model patterns.

False Positive Prevention

Critical requirements:

• Minimum 200 words for reliable analysis
• Never flag on single indicators alone
• Use ensemble scoring (multiple signals required)

High false-positive risk groups:

• Non-native English speakers (61% false positive rate in research)
• Technical/formal writing
• Neurodivergent writers
• Content using grammar correction tools

Ineffective indicators (do NOT rely on these):

• Perfect grammar alone
• "Bland" or "robotic" prose
• "Fancy" or unusual vocabulary
• Letter-like formatting alone
• Conjunctions starting sentences

Signs of human writing:

• Text from before November 30, 2022
• Ability to explain editorial choices
• Personal anecdotes with verifiable details
• Minor errors and natural quirks

See reference/false-positive-prevention.md for detailed guidance.

Analysis Output Format

Structure findings as:

**Overall Assessment**: [Likely AI / Possibly AI / Likely Human / Inconclusive]
**Confidence**: [Low / Medium / High]

**Summary**: 2-3 sentence overview

**Evidence Found**:
- [Category]: [Specific indicator] - "[Quote from text]"
- [Category]: [Specific indicator] - "[Quote from text]"

**Mitigating Factors**: [Elements suggesting human authorship]

**Caveats**: [Limitations, alternative explanations]

Key Principles

1. No certainty claims - AI detection is probabilistic
2. Multiple signals required - Single indicators prove nothing
3. Context matters - Academic writing differs from blogs
4. Stakes awareness - False accusations cause real harm
5. Evolving field - Detection methods require constant updates

Reference Files

• vocabulary-patterns.md - Complete word/phrase lists with frequencies
• structural-patterns.md - Sentence, paragraph, and discourse patterns
• content-patterns.md - Importance puffery, promotional language, content tells
• formatting-patterns.md - Title case, boldface, emojis, visual patterns
• markup-artifacts.md - Technical artifacts: turn0search, oaicite, Markdown, tracking
• citation-patterns.md - Broken links, invalid identifiers, hallucinated references
• model-fingerprints.md - GPT, Claude, Gemini, Grok, Perplexity specific tells
• false-positive-prevention.md - Avoiding false accusations, ineffective indicators

Sources

This knowledge base synthesizes research from:

• Stanford HAI (DetectGPT, bias studies)
• GPTZero, Originality.ai, Turnitin, Pangram methodologies
• Academic papers on stylometry and discourse analysis
• Empirical studies on detection accuracy and limitations
• Wikipedia:WikiProject AI Cleanup field guide (2025)
• Community-documented patterns from Wikipedia editing