wikipedia-not-ai

Detect AI-generated text using Wikipedia-based patterns and structural signals.

AI Text Detection Guide

Based on Wikipedia's community-maintained guide (editorial heuristics, not a definitive detector). Applicable to any text — not just Wikipedia.

How to analyze text

Scan for signals across the categories below.
Weight clusters more heavily than isolated occurrences:

  • One signal → coincidence
  • Multiple signals in proximity → strong indicator

Output format (strict)

For each detected signal:

  • Quote: exact fragment from the text
  • Category: which signal type it belongs to
  • Explanation: short, specific reasoning

Then:

Final verdict: Likely AI / Possibly AI / Likely Human
Confidence: Low / Medium / High

1. AI Vocabulary Density (strongest signal)

LLMs statistically overuse specific words. Check for density and clustering.

See word-lists.md for full lists.

Quick scan examples:
pivotal, underscore, tapestry, delve, meticulous, vibrant, intricate, testament, bolstered, garner, fostering, showcasing, align with, enhance, highlighting

Rule:

  • 1 word → ignore
  • 3+ → suspicious
  • 5+ in short passage → strong signal

2. Content Patterns

Significance inflation

Adds artificial importance to neutral facts.

Examples:

  • marks a pivotal moment
  • represents a shift
  • underscores its importance
  • reflects broader trends

Challenges formula

Rigid structure:

  • "Despite X, it faces challenges..."
  • "Despite these challenges..."

Superficial -ing clauses

Adds fake analytical depth at sentence end:

  • highlighting its importance
  • underscoring the significance

Vague attribution (weasel language)

No clear source:

  • Experts argue
  • Studies suggest
  • Observers have noted

Promotional / travel tone

Unnaturally warm or рекламный стиль:

  • vibrant
  • breathtaking
  • rich cultural heritage
  • renowned

3. Sentence Structure Tells

Copula avoidance

Avoids simple verbs:

  • ❌ serves as → ✅ is
  • ❌ boasts → ✅ has

Negative parallelisms

Formulaic balance:

  • Not only X, but also Y
  • It's not just X, it's Y

Rule of three

Unnatural triplets:

  • X, Y, and Z used excessively

Elegant variation

Unnatural synonym switching:

  • subject → the figure → this individual

4. Formatting Tells

  • Title Case headings everywhere
  • Excessive bold text
  • Emoji usage 🎯
  • Em dash overuse
  • Inline bold definition lists
  • Unnecessary tables
  • Curly quotes

5. Communication Leakage

Artifacts from chat interface:

  • I hope this helps
  • Let me know if you need
  • Would you like me to

Also:

  • placeholders [INSERT...]
  • subject lines inside body
  • meta-statements about the document

6. Citation & Markup Artifacts

Technical residue:

  • turn0search0
  • contentReference[...]
  • oai_citation
  • utm_source=chatgpt.com
  • [attached_file:1]
  • <grok_card>

7. Older Model Tells (2022–2024)

  • It's important to note that...
  • In conclusion...
  • Overall...
  • abrupt cutoffs

False Positives — do NOT flag alone

  • Correct grammar
  • Formal tone
  • Transition words
  • Long structured text
  • Unsourced claims

Scoring guideline

  • 1–2 weak signals → Possibly AI (low confidence)
  • 3–5 mixed signals → Likely AI (medium confidence)
  • 5+ strong clustered signals → Likely AI (high confidence)

Clusters matter more than isolated signals.

Reference