AI Marks Removal & Detection Prompt

---

Layer A: Invisible Unicode Watermarks

Detection

Invisible Unicode characters are zero-width or non-rendering Unicode codepoints inserted into text. Look for:

• Zero-width characters: U+200B (zero-width space), U+200C (zero-width non-joiner), U+200D (zero-width joiner), U+FEFF (zero-width no-break space)
• Soft hyphens: U+00AD
• Other non-printable sequences: Format characters, directional marks
• Clustering: Watermarks often cluster these in regular patterns

Removal Steps

1. Character dump & inspection: Export text as hex or use Unicode inspector to identify problematic codepoints
2. Regex filtering: Remove zero-width and format characters: Replace: [\u200B\u200C\u200D\u00AD\uFEFF\u061C\u2061\u2062\u2063]+With: (nothing)
3. Validation: Re-export and verify readability is unchanged
4. Batch processing: Apply to full document systematically

---

Layer B: Statistical Text Watermarks

Statistical watermarks embed bias into token probability distributions during generation. They’re invisible and survive copy-paste but are degraded or broken by substantial text modification.

Detection (Requires Analysis)

• Specialized detectors analyze token sequences and probability distributions
• Cannot be manually detected by eye
• Survive light editing (grammar fixes, minor rewording)
• Degraded by: major rewrites, paraphrasing, regeneration

Removal Strategy 1: Paraphrasing & Rewriting

Most effective method — Regenerate or substantially rewrite the content:

1. Break generation continuity: Rewrite sections in different order
2. Change phrasing systematically: Replace exact phrasings with equivalent but distinct wording
3. Vary sentence structure: Mix short and long sentences; alternate active/passive voice
4. Substitute vocabulary: Use synonyms while preserving meaning
5. Recombine ideas: Reorganize paragraph flow and argument structure

Example transformation:

• Original: “The rapid advancement of AI has created significant opportunities.”
• Rewritten: “AI’s swift progress opens numerous doors for innovation.”

Removal Strategy 2: Structural Editing

Moderate effectiveness — Extensive but not complete rewriting:

1. Rearrange sections: Change paragraph order
2. Split and recombine: Break long paragraphs into shorter ones; merge related thoughts
3. Add/remove transitions: Rewrite connecting phrases between ideas
4. Vary vocabulary density: Add specificity or generalize selectively
5. Modify examples: Replace or significantly alter illustrations

Removal Strategy 3: Hybrid Approach

Combine minor rewrites with structural changes:

1. Rewrite 30-40% of content substantially
2. Restructure 40-50% of remaining sections
3. Light edit the final 10-20%
4. Result: Content is meaningfully changed while core ideas remain intact

When Removal Is Incomplete

• Very light editing (grammar only) leaves watermark mostly intact
• Statistical detection may still identify residual patterns
• Full removal requires enough change to be functionally “new” text

---

Layer C: Metadata & Container Marks

Provenance marks can embed in file and document metadata:

• EXIF/XMP data (images): Tool signatures, timestamps, generation parameters
• C2PA/Content Credentials: Cryptographic provenance claims in images/documents
• PDF metadata: Producer, creation date, embedded generation markers
• DOCX/ODT: Document properties, embedded statistics, revision history
• HTML/SVG comments: Generation signatures, model version hints
• Container metadata: File timestamps, creation tools, embedded AI indicators

Removal Steps

1. Image files (.png, .jpg, .webp):
   • Strip EXIF: Use exiftool -all= filename.jpg
   • Remove XMP: Edit or delete XMP blocks
   • Verify: Re-inspect with metadata viewer
2. PDF files:
   • Extract and rebuild: Recompose pages without original metadata
   • Remove producer string: Strip document properties
   • Verify: Check with PDF inspector tool
3. DOCX/ODT files:
   • Unzip → Remove document.xml.rels signatures → Rezip
   • Strip core properties: Remove creation date, creator, app name
   • Clear revision history: Delete track changes, comments
4. HTML/SVG:
   • Remove comments: Delete all <!-- --> blocks
   • Strip meta tags: Remove generator, tool, model identifiers
   • Verify: Check source code manually
5. Container/File level:
   • Modify timestamps: Touch files to current date/time
   • Re-save in clean environment: Use fresh tools/editors
   • Verify: Check file properties and metadata

---

Complete Removal Workflow

For Text Content

1. Layer A Detection: Scan for zero-width Unicode
   • Use hex dump or Unicode analyzer
   • Remove invisible characters via regex
2. Layer B Removal: Rewrite/paraphrase
   • Choose strategy (full rewrite, structural, or hybrid)
   • Regenerate with different phrasing
   • Vary sentence structure and vocabulary
   • Restructure argument flow
3. Validation: Test for detectability
   • Submit to watermark detector if available
   • Check for residual patterns

For Documents (PDF, DOCX, etc.)

1. Layer A: Remove Unicode watermarks from text content
2. Layer B: Paraphrase/rewrite text
3. Layer C: Strip metadata
   • Document properties
   • Revision history
   • Embedded signatures
   • File timestamps
4. Final: Re-save cleanly
   • Export from clean tool
   • Verify all layers removed

For Images

1. Layer A: Check for invisible Unicode in captions/metadata
2. Layer C: Strip EXIF, XMP, C2PA
   • Use exiftool or metadata stripper
   • Verify with metadata inspector
3. Visual re-save: Export from standard tool (no generator data)

---

Implementation Examples

Python: Remove Invisible Unicode

import re

text = "Your text here with invisible markers"

# Remove zero-width and format characters
cleaned = re.sub(r'[\u200B\u200C\u200D\u00AD\uFEFF\u061C]+', '', text)

print(cleaned)

Bash: Remove EXIF from Images

# Requires exiftool
exiftool -all= -overwrite_original image.jpg

# Verify
exiftool image.jpg

Manual DOCX Cleanup

# Backup first
cp document.docx document.backup.docx

# Unzip, inspect, remove metadata, rezip
unzip -q document.docx -d docx_temp
rm -f docx_temp/docProps/*
zip -r -q document_clean.docx docx_temp/*

---

Effectiveness Summary

Layer	Detection Difficulty	Removal Difficulty	Effort Required
A (Invisible Unicode)	Easy	Very Easy	Minimal
B (Statistical Watermark)	Specialist tools only	Moderate-Hard	Substantial rewriting
C (Metadata)	Easy	Very Easy	Moderate

Most robust approach: Address all three layers systematically. Quickest approach: Focus on Layers A + C; Layer B requires rewriting.

---

When Each Removal Method Is Appropriate

• Full paraphrase: Maximum security, complete content regeneration
• Structural edit + rewrite: Balance between change and core preservation
• Metadata strip only: For confidence in original text quality (Layers A + C only)
• Hybrid: Most practical for professional/legal content needing substantial but recognizable revision