AI Watermark Remover

AI Text Watermark Remover Skill (Claude, Gemini, SynthID & GPT)

This skill provides an end-to-end operational protocol to detect and eliminate all forms of AI text watermarking across Claude (Anthropic), Gemini & SynthID-Text (Google DeepMind), and GPT (OpenAI).

It targets three distinct watermarking layers:

1. Invisible Steganography (Zero-width characters, homoglyphs, invisible separators).
2. Statistical / $n$-gram Logit Watermarks (Google SynthID-Text, Kirchenbauer green/red list, Aaronson Gumbel watermarks).
3. Model-Specific Stylistic Fingerprints (Claude-isms, Gemini tropes, predictable cadence, and low-entropy formatting).

1. The Anatomy of AI Text Watermarks

Modern AI text generators (including Claude, GPT, and Gemini) leave traces across three distinct vectors:

2. The 5-Stage Dewatermarking Protocol

When given text to dewatermark, execute the following 5 stages sequentially:

[Input Text]
     │
     ▼
┌────────────────────────────────────────────────────────┐
│ Stage 1: Deterministic Unicode & Binary Sanitization   │ ➔ Strip zero-width chars & normalize NFKC
└────────────────────────────────────────────────────────┘
     │
     ▼
┌────────────────────────────────────────────────────────┐
│ Stage 2: Statistical $n$-gram Chain Shattering          │ ➔ Reorder clauses & switch voice (breaks SynthID)
└────────────────────────────────────────────────────────┘
     │
     ▼
┌────────────────────────────────────────────────────────┐
│ Stage 3: Lexical & Stylistic "De-Claude-ification"     │ ➔ Replace 50+ AI clichés & strip hedging
└────────────────────────────────────────────────────────┘
     │
     ▼
┌────────────────────────────────────────────────────────┐
│ Stage 4: Burstiness & Perplexity Calibration           │ ➔ Inject rhythm variation & natural idioms
└────────────────────────────────────────────────────────┘
     │
     ▼
┌────────────────────────────────────────────────────────┐
│ Stage 5: Semantic & Fact-Checking Preservation         │ ➔ Verify 0% hallucination & 100% intent retained
└────────────────────────────────────────────────────────┘
     │
     ▼
[100% Clean, Humanized Text]

Stage 1: Deterministic Unicode & Binary Sanitization

First, strip all invisible characters and normalize Unicode.

Regex to strip: [\u200B\u200C\u200D\u200E\u200F\u202A-\u202E\u2060-\u206F\uFEFF\u00AD\u180E\u034F\u115F\u1160\u3164\uFFA0]

(You can run python3 scripts/sanitize.py "<text>" or execute the regex directly.)

Stage 2: Statistical $n$-gram Chain Shattering

Statistical watermarks (like SynthID) hash previous words $w_{t-k} \dots w_{t-1}$ to determine the pseudo-random score of token $w_t$. Disrupt this sequence:

1. Clause Inversion: Shift dependent clauses from the end of the sentence to the beginning, or vice versa.
   • AI: "Machine learning improves precision because it optimizes loss gradients."
   • Clean: "By optimizing loss gradients, machine learning delivers sharper precision."
2. Voice Shifting: Convert passive constructions into active voice with explicit agency, or split long compound sentences into crisp independent clauses.
3. Punctuation & Connector Alteration: Replace predictable transitions ("Furthermore", "Moreover") with contextual connectors ("On top of that", "In practice", "Similarly", or simple conjunctions like "And" / "But").

Stage 3: Lexical & Stylistic "De-Claude & De-Gemini-ification"

Both Claude (Anthropic) and Gemini (Google) have recognizable vocabulary biases and structural mannerisms. Replace them systematically:

Banned AI Buzzwords & Substitutions:

• ❌ delve / delve into / dive deep ➔ investigate, explore, look at, examine, break down
• ❌ rich tapestry / tapestry / treasure trove ➔ mix, collection, blend, ecosystem, landscape
• ❌ testament to / stands as a ➔ proof of, evidence that, shows, reflects
• ❌ multifaceted / nuanced / plethora / myriad ➔ complex, layered, detailed, varied, wide range
• ❌ beacon / beacon of / powerhouse ➔ leader in, example of, model for, strong platform
• ❌ foster / fostering / supercharge ➔ encourage, build, support, grow, speed up, drive
• ❌ underscores / highlights / shines a light ➔ shows, stresses, makes clear, points to
• ❌ pivotal / paramount / crucial / critical role ➔ key, vital, essential, important, direct factor
• ❌ game-changer / revolutionizes / transforms ➔ major upgrade, big step forward, improves
• ❌ harness the power of / unleash ➔ use, apply, run, deploy, leverage
• ❌ elevate / supercharge your workflow ➔ improve, speed up, streamline, refine
• ❌ intertwined / seamlessly ➔ connected, combined, directly, smoothly
• ❌ certainly! / sure thing! / here is a breakdown ➔ (Delete completely; start directly with the answer)
• ❌ let's unpack this / let's dive in ➔ (Delete completely; begin immediately)
• ❌ in conclusion / to sum up ➔ (Delete or use contextual closure: "Ultimately", "At its core")
• ❌ it is important to remember / keep in mind that ➔ (Delete; state the point directly)

Banned Structural Tropes:

1. The "Symmetrical Triad": AI frequently creates lists of exactly 3 bullet points with bold headers of identical length. Fix: Use varying list lengths (2, 4, or 5 items) or narrative paragraphs.
2. The "Balanced Both-Sides Hedge": "While X has merits, it is equally important to acknowledge Y." Fix: Take a clear, decisive stance or state the tradeoff straightforwardly.
3. The Em-Dash Overuse: AI overuses — to append decorative thoughts. Fix: Use parentheses, commas, or start a new sentence.
4. The "Bold Colon" Formatting: Gemini & Claude heavily output **Feature:** Explanation. Fix: Integrate descriptions directly into natural running sentences.

Stage 4: Burstiness & Perplexity Calibration

Human writing is asymmetrical and irregular. AI writing is overly uniform and predictable.

• Sentence Length Burstiness:
  • Write a 4-word sentence.
  • Follow it with an 18-word explanation providing technical context.
  • Follow that with a medium 10-word summary.
  • Target a sentence length standard deviation $\sigma > 8.0$.
• Vocabulary Perplexity:
  • Use specific, domain-exact nouns rather than broad conceptual abstractions.
  • Use colloquial phrasing where appropriate for tone (e.g., "cut down" instead of "drastically mitigate").

Stage 5: Semantic & Fact-Checking Preservation

• Ensure no technical terms, metrics, numerical values, or citations are lost or altered.
• Verify that tone matches the user's intent (Academic, Technical, Business, or Casual).

3. Multi-Domain Persona Presets

Select the target persona to match the required domain:

🎓 1. Academic & Scholarly Preset

• Goal: Pass Turnitin / GPTZero while retaining scientific rigor and exact citations.
• Rules:
  • Use precise, peer-reviewed domain terminology; avoid colloquialisms.
  • Eliminate rhetorical questions and filler ("It is important to note", "In conclusion").
  • Use active academic voice ("The data indicates" rather than "It can be observed from the data that").
  • Maintain variable paragraph densities and asymmetrical citation groupings.

💻 2. Software Engineering & Technical Preset

• Goal: Direct, pragmatic technical documentation with 100% code syntax immunity.
• Rules:
  • IMMUNITY SHIELD: Never alter code inside ``` blocks, inline backticks `code`, or terminal commands.
  • Strip promotional adjectives ("revolutionary architecture", "seamless scalability").
  • Explain mechanisms step-by-step with concrete inputs, outputs, and failure modes.

💼 3. Executive & Business Preset

• Goal: Crisp, action-oriented, data-first corporate communication.
• Rules:
  • Lead with bottom-line outcomes, metrics, percentages, and deadlines.
  • Drop conceptual fluff ("delve into the landscape", "harness the power").
  • Use asymmetrical bullet points (e.g., 2 or 4 items, varying in length and structure).

✍️ 4. Conversational & Creative Preset

• Goal: Authentic human storytelling with genuine voice and cadence.
• Rules:
  • Use natural contractions ("didn't", "won't", "we've"), idioms, and rhetorical pauses.
  • Break rhythmic predictability by placing 3-word punchy sentences after detailed observations.
  • Avoid formulaic summaries at the end.

4. Multi-Lingual Anti-Watermark Rules

AI watermarks appear across international languages. Apply these translation-specific bans:

• 🇪🇸 Spanish: Ban "es fundamental destacar", "un tapiz de", "en conclusión", "un papel crucial", "desempeña un papel", "un faro de".
• 🇫🇷 French: Ban "il convient de noter", "un rôle primordial", "témoignage de", "en conclusion", "un éventail de", "un phare de".
• 🇩🇪 German: Ban "es ist wichtig zu beachten", "ein facettenreicher", "zusammenfassend lässt sich sagen", "eine entscheidende rolle", "ein meilenstein".

5. Universal Dewatermarking Prompts

General Dewatermarking Prompt

Rewrite the following text using the ai-watermark-remover protocol.
Domain Persona: [Academic / Technical / Business / Casual]

Strict Rules:
1. Strip all AI buzzwords (delve, tapestry, testament, multifaceted, foster, beacon, nuanced, underscores, paramount, crucial role, harness the power, supercharge, let's unpack).
2. Shatter n-gram sequences (SynthID) by inverting clauses and switching between active and passive forms.
3. Inject burstiness: mix punchy 4-word sentences with longer 20+ word compound sentences.
4. Protect all code blocks, syntax, and math equations ($...$) completely intact.
5. Make the tone authentic, authoritative, and direct.

Text to dewatermark:
"""
[PASTE TEXT HERE]
"""

6. Testing & Automated Scorecard

Run the offline detector simulator on any text or file:

# Run heuristic scorecard (Human %, Burstiness, SynthID risk, Markers)
python3 scripts/sanitize.py "<text or file_path>"

# Output JSON for pipeline integration
python3 scripts/sanitize.py input.md --json

# Run real-time clipboard monitor
python3 scripts/clipboard_daemon.py

• Detection Checkers (GPTZero / Turnitin / CopyLeaks): < 10% AI Probability