anti-ai-validator

Anti-AI Validator

Quick Reference

Purpose

The Anti-AI Validator systematically detects and removes AI-generated text patterns to ensure natural, human-level quality Korean writing.

Detection Categories

Critical Pattern Detection:

• "~를 넘어 ~하다" construction (top priority)
• Forbidden vocabulary from the overused terms list
• Speculative expressions and evasive language
• Repetitive word usage within paragraphs

Structural Detection:

• Predictable sentence patterns
• Uniform sentence length distribution
• Limited connector diversity
• Abstract metaphorical constructions

Vocabulary Detection:

• Word repetition within paragraphs
• Hollow emphasis words and filler words
• Overused technical buzzwords
• Generic transition phrases

Verification Process

The validator runs a 4-step scan process:

Step 1 - Critical Pattern Scan:

• Search for all forms of the "~를 넘어" pattern
• Immediately flag all instances for replacement
• No exceptions allowed

Step 2 - Vocabulary Scan:

• Compare content against the forbidden word list
• Identify repetitions within paragraph boundaries
• Flag hollow emphasis words and filler words

Step 3 - Structure Scan:

• Analyze sentence length distribution
• Check connector diversity
• Evaluate paragraph variation

Step 4 - Generate Corrections:

• Provide specific replacements for each violation
• Suggest alternative constructions
• Recommend vocabulary rotation

Implementation Guide

Critical Pattern: "~를 넘어" Detection

This pattern is the most easily detected marker of AI writing in Korean. The validator flags all instances without exception.

Detection Rules:

• Scan for "넘어" preceding an action verb or noun
• Include all variants: "를 넘어", "을 넘어", "넘어서"
• Flag even when separated by other words

Common Violations:

• "단순한 응답자를 넘어 능동적인 문제 해결자로서"
• "코드 생성을 넘어 전체 프로젝트를 자동화한다"
• "기본 기능을 넘어 고급 분석까지 제공한다"

Approved Replacements:

• "~에 그치지 않고 ~도 수행한다"
• "~뿐만 아니라 ~도 담당한다"
• "~뿐 아니라 ~까지 포함한다"
• "~와 더불어 ~도 제공한다"
• "~은 물론, ~에도 영향을 미친다"

Forbidden Vocabulary Detection

The validator maintains a comprehensive list of forbidden terms categorized by type.

Overused abstract terms:

• "paradigm" → "approach", "methodology", "framework"
• "innovation" → "new feature", "improved feature", "enhanced approach"
• "game-changer" → "important tool", "effective tool"
• "groundbreaking" → "notable", "substantial", "state-of-the-art"
• "amazing" → "impressive", "effective", "practical"

Empty emphasis words:

• "very" → replace with a specific quantity or measurement
• "quite" → replace with specific data or benchmarks
• "really" → replace with factual evidence
• "of course", "naturally" → remove entirely

Filler expressions:

• "significantly" → replace with a specific percentage or metric
• "considerably" → replace with comparative data
• "clearly" → replace with "was confirmed by ~"
• "certainly" → replace with "according to ~"
• "obviously" → replace with an evidence-based statement

Speculative expression detection

Speculative language weakens credibility and signals AI generation. The validator removes all speculative patterns.

Prohibited endings:

• "will likely be ~" in all contexts
• "is expected to ~"
• "appears to ~"
• "may ~" (polite speculation)

Prohibited words:

• "perhaps", "maybe", "possibly"

Required replacements:

• Declarative endings: "~does", "~provides", "~supports"
• Confirmed facts: "was confirmed by ~", "according to ~"
• Data-based statements: "based on ~", "on the basis of ~"

Conversion examples:

• Bad: "Claude Code is expected to soon support a new programming language"
• Good: "As of November 2025, Claude Code supports more than 20 languages, including Python, JavaScript, and TypeScript (1)"
• Bad: "This feature may improve productivity"
• Good: "This feature improves productivity by an average of 35% (2)"

Vocabulary repetition detection

The rule of three strengthens lexical diversity:

Within-paragraph rules:

• The same word does not appear more than once within a single paragraph
• A third use requires synonym substitution
• Track word usage across paragraph boundaries

Within-section rules:

• The same phrase is not repeated more than once within a section
• Find an alternative phrase for the second use

Synonym rotation system:

High-frequency verbs:

• "provides" rotation: "supports", "supplies", "presents", "offers"
• "handles" rotation: "deals with", "manages", "executes", "performs"
• "uses" rotation: "utilizes", "applies", "adopts"

High-frequency nouns:

• "Function" cycle: "role", "responsibility", "task", "capability"
• "System" cycle: "system", "structure", "platform", "environment"
• "Method" cycle: "method", "technique", "procedure", "strategy"

Sentence structure analysis

AI-generated text shows predictable syntactic patterns. The validator checks for structural diversity.

Active/passive alternation:

• Active example: "The system processes data"
• Passive example: "Data is processed by the system"
• Mixed example: "Data is processed by the system, and results are generated through this"

Length distribution:

• Short sentence (15-20 words): directly conveys the core concept
• Medium sentence (25-35 words): provides necessary context and explanation
• Long sentence (40-50 words): handles complex relationships between concepts
• Consecutive sentences in the same length category flag

Connective distribution:

• Sequential connectives: "and", "also", "not only that"
• Contrastive connectives: "however", "on the other hand", "nevertheless"
• Causal connectives: "so", "therefore", "accordingly", "in conclusion"
• Transitional connectives: "meanwhile", "on the other hand", "also"
• Overuse of a single connective type flag

Abstract metaphor detection

Vague metaphorical patterns obscure meaning and signal AI generation. The validator replaces them with direct explanations.

Prohibited metaphor patterns:

• Abstract A-B pattern: "in a pattern where B expands on the foundation of A"
• Forced structural explanation: "in a cause-effect structure called ~"
• Music/art metaphors for technical concepts: "like a symphony, multiple elements are in harmony"

Required direct explanations:

• Causality: "as A occurred, B appeared"
• Functionality: "A directly performs ~, and through ~ creates concrete results"
• Comparison: "unlike A, B shows the following differences"

Transformation examples:

• Bad: "Technological development proceeds in a pattern where B is further expanded on the foundation of A"
• Good: "Technological development further expanded new functionality B based on existing technology A"
• Bad: "This is a structure of change over time, where in the past it was A and now it is B"
• Good: "Over time, the technology developed from A to B"

Advanced patterns

Validation report generation

The validator generates a detailed report for content review:

Violation summary:

• Total number of violations by category
• Critical vs non-critical classification
• Ranking of correction priorities

Violation details:

• Exact text location (paragraph number, sentence position)
• Specific violation type
• Recommended replacement
• Rationale for change

Revision Guidance:

• Step-by-step revision instructions
• Alternative phrasing for each violation
• Examples of appropriate usage

Automatic revision suggestions

The validator provides context-aware replacement suggestions:

Key patterns:

• Immediate replacement from the list of approved alternatives
• Multiple options ranked by appropriateness for the context
• Usage examples demonstrating correct application

Vocabulary violations:

• Synonym suggestions in a recursive system
• Context preservation guidelines
• Natural flow verification

Structural issues:

• Sentence restructuring recommendations
• Length adjustment strategies
• Alternative connector options

Self-revision process

When a violation is detected, follow this revision workflow:

Step 1 - Identify the violation type:

• Match the violation with a category in this document
• Determine severity (critical, high, medium, low)

Step 2 - Select an appropriate replacement:

• Choose from approved alternatives
• Consider the context and surrounding sentences

Step 3 - Rewrite the sentence:

• Apply the replacement while maintaining natural flow
• Check that no new violations have been introduced

Step 4 - Check the surrounding context:

• Check that no new repetition issues have been created
• Verify that connector usage is still appropriate
• Check that paragraph cohesion is maintained

Step 5 - Read-aloud test:

• Read the paragraph aloud for a natural Korean rhythm
• Check that no awkward expressions have been introduced
• Check that readability has improved

Things that work well together

Skills:

• writing-standards - provides the broader quality framework that the validator enforces
• personas - verifies that persona-specific content meets quality standards
• research - checks that citations and quoted sources follow standards

Commands:

• post - final content verification before publishing

Verification checklist

Pre-publish verification

Pattern checks:

• No "beyond ~" pattern anywhere in the text
• Zero instances of prohibited vocabulary
• No speculative endings or evasive language
• No repetition of the same word twice within a single paragraph

Structure checks:

• Sentence length varies appropriately (no three or more consecutive sentences in the same category)
• Connector types are appropriately distributed
• No abstract metaphorical phrases

Revision checks:

• All flagged violations have been addressed
• Replacement suggestions have been correctly implemented
• Natural flow is maintained after revision
• No new violations introduced during revision

Final quality check:

• Passed read-aloud test for natural rhythm
• Technical accuracy preserved
• Readability improved from the original
• Human-level quality writing achieved

---

Version: 2.0.0 Last Updated: 2026-01-06 Category: Quality Assurance Type: Content Validation Status: Production Ready