ai-text-audit 1.0.0

Detect AI writing patterns. Make text human again.

ai-text-audit 🦞

Detect AI writing patterns. Make text human again.

A fast, offline CLI tool and Python library that detects AI-generated text patterns based on linguistic analysis — no API keys, no cloud, no privacy concerns.

Built from analysis of 10,000+ AI-generated texts, based on Wikipedia's comprehensive guide to AI writing patterns.

✨ Features

• 🔍 22+ AI pattern detectors — em-dashes, rule of three, AI vocabulary, promotional language, vague attribution, and more
• 📊 Weighted scoring — patterns ranked by severity, not just frequency
• 🎯 Smart suggestions — actionable rewrite tips for each detected pattern
• 🔄 Pipeline mode — --json output for CI/CD integration
• ⚡ Fast — analyzes 1000 words in <50ms, no network calls
• 🌍 Offline — zero external dependencies for core detection
• 🛠️ Extensible — add your own patterns with simple YAML config
• 📝 Batch mode — process entire directories or stdin

🚀 Quick Start

pip install ai-text-audit

# Analyze a file
ai-audit README.md

# Analyze from stdin
echo "This groundbreaking tool represents a pivotal advancement..." | ai-audit -

# JSON output for pipelines
ai-audit --json document.md

📖 Usage

CLI

# Basic analysis
ai-audit document.txt

# With detailed pattern breakdown
ai-audit --verbose document.txt

# JSON output
ai-audit --json document.txt > report.json

# Set threshold (0-100, default: 30)
ai-audit --threshold 50 document.txt

# Show only score (for scripting)
ai-audit --score-only document.txt

# Batch directory
ai-audit --batch ./docs/

# Pipeline: check then humanize
cat draft.md | ai-audit - --humanize

Python API

from ai_text_audit import Auditor

auditor = Auditor()
result = auditor.analyze("This groundbreaking innovation represents a pivotal moment...")

print(result.score)          # 72.3 (0-100, higher = more AI-like)
print(result.verdict)        # "likely_ai" | "possibly_ai" | "likely_human"
print(result.patterns)       # List of detected patterns with counts
print(result.suggestions)    # Rewrite suggestions per pattern

# Custom patterns
auditor = Auditor(patterns_file="my_patterns.yaml")

# Batch
results = auditor.analyze_dir("./docs/")

Output Examples

Terminal (default):

🦞 AI Text Audit — document.md
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Score: 72.3 / 100  →  🤖 Likely AI

Patterns detected:
  ⚠️  em_dash (×8)              "—" appears frequently, uncommon in casual writing
  ⚠️  ai_vocab (×12)            Words like "crucial", "pivotal", "tapestry"
  ⚡ rule_of_three (×5)          "X, Y, and Z" structure overused
  ⚡ boast_language (×3)         "stands as a testament", "boasts"
  💡 promotional (×2)           "groundbreaking", "cutting-edge"

Suggestions:
  → Replace "—" with ", " or "."
  → Swap "crucial" → "important", "pivotal" → "key"
  → Vary list structures beyond "A, B, and C"

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

JSON (--json):

{
  "file": "document.md",
  "score": 72.3,
  "verdict": "likely_ai",
  "word_count": 847,
  "patterns": [
    {"name": "em_dash", "count": 8, "weight": 1, "severity": "low", "description": "..."},
    {"name": "ai_vocab", "count": 12, "weight": 1, "severity": "medium"}
  ],
  "suggestions": ["Replace em-dashes with commas..."]
}

🔧 Detection Patterns

Pattern	Weight	Example	Why It's AI
em_dash	1	—	Overused in AI, rare in casual writing
ai_vocab	1	"crucial", "tapestry", "delve"	Hallmark AI vocabulary
rule_of_three	0.5	"A, B, and C"	AI loves triplets
negative_parallelism	2	"It's not just X, it's Y"	Classic AI structure
boast_language	2	"stands as a testament"	Promotional framing
promotional	2	"groundbreaking", "revolutionary"	Buzzword density
vague_attribution	2	"experts say", "many believe"	Fake authority
collaborative_artifact	3	"I hope this helps!"	Chatbot leftovers
filler	1	"It is important to note"	Academic filler phrases
ing_superficial	2	"...ing, reflecting"	Superficial analysis pattern

See docs/patterns.md for the full list of 22+ patterns.

🛠️ Configuration

Create ~/.ai-audit.yaml or pass --config:

# Custom patterns
patterns:
  my_custom:
    regex: "\\b(synergy|leverage|optimize)\\b"
    weight: 1
    description: "Corporate buzzwords"

# Thresholds
thresholds:
  likely_ai: 60
  possibly_ai: 30

# Ignore files
ignore:
  - "*.min.js"
  - "vendor/**"

🏗️ CI/CD Integration

GitHub Actions

- name: Check for AI writing
  run: |
    pip install ai-text-audit
    ai-audit --threshold 50 --json . > audit.json
    SCORE=$(jq '.score' audit.json)
    if (( $(echo "$SCORE > 50" | bc -l) )); then
      echo "::warning::Content score is $SCORE — likely AI-generated"
    fi

Pre-commit Hook

repos:
  - repo: https://github.com/sudabg/ai-text-audit
    rev: v1.0.0
    hooks:
      - id: ai-audit
        args: ['--threshold', '40']

📦 Installation

# PyPI (recommended)
pip install ai-text-audit

# From source
git clone https://github.com/sudabg/ai-text-audit.git
cd ai-text-audit
pip install -e .

# With optional dependencies
pip install ai-text-audit[full]  # includes rich, pyyaml

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md.

Good first issues:

• Add new detection patterns
• Improve existing pattern accuracy
• Add support for more languages (currently optimized for English + Chinese)
• Write tests for edge cases

📄 License

MIT License — see LICENSE.

🙏 Acknowledgments

• Pattern analysis based on Wikipedia's Signs of AI Writing
• Inspired by the need for transparent, offline AI content detection
• Built with 🦞 by 小哩子

📈 Roadmap

• v1.0.0 — Core detection + CLI (this release)
• v1.1.0 — Stylometric analysis (sentence length variance, vocabulary richness)
• v1.2.0 — Perplexity-based detection (lightweight, local)
• v1.3.0 — Multilingual support (Japanese, Korean, etc.)
• v2.0.0 — Self-hosted API server mode