How to Remove AI Watermarks From Text: What Works and What Doesn't (2026)
16 min read
The way to remove an AI watermark from text depends on which of three kinds you have. Hidden characters, the AI markers that copy-paste leaves behind, come out with find and replace or a text cleaner. File metadata drops off when you paste the words into a fresh document. Statistical AI signatures from Gemini, ChatGPT and Claude are part of the word choice, so only a real rewrite changes them.
Provider watermarks became far more common in 2026. Supported Claude models have watermarked text since August 2026, and OpenAI began watermarking ChatGPT text for EU users in October 2026.
Below are step-by-step fixes for Microsoft Word and Google Docs, plus an honest look at what no edit can remove.
Clean Hidden AI Characters With Phrasly's AI Text Watermark Remover 👇
What Is an AI Text Watermark? Markers, Metadata and AI Signatures in 2026

An AI text watermark is any signal that marks writing as machine-made. In 2026, three different things go by that name: hidden Unicode characters inside the text, metadata attached to the file, and statistical signatures that Gemini, ChatGPT and Claude build into their word choice. Each one needs its own fix, and only the first two can be deleted.
This article focuses on text watermarks, not image watermarks.
AI detectors such as GPTZero and Originality.ai are a separate thing, because they don't add or read watermarks. They detect text based on patterns in predictability and phrasing associated with LLM-written text.
A true statistical watermark is SynthID Text by Google DeepMind. It is a keyed statistical watermark, which means a secret key nudges which words the model picks.
SynthID-Text embeds a signal directly into the tokens (words) that a model chooses as it generates. Using token probability watermarking, SynthID creates a watermark associated with AI content provenance (the origin of content, not ownership).
Type | What it is | Where you find it | Survives copy-paste? | How to remove it |
|---|---|---|---|---|
Hidden characters (AI markers, AI tags) | Zero-width spaces, non-breaking and narrow no-break spaces, soft hyphens | Inside the text | ✅ Yes | Find and replace in Word or Google Docs, or a text cleaner |
File metadata | Author, edit time, app name; signed C2PA credentials on some AI-made files | On the file, not the words | ❌ No, it stays with the file | Paste into a new document; Word Document Inspector |
Statistical watermark (AI signature) | A secret-key pattern in word choice (SynthID Text, textGrain, Claude) | In the words themselves | ✅ Yes | No character to delete; vendors say heavy rewriting or translation weakens it |
Most hidden characters are copy-paste and formatting leftovers, not official watermarks. In April 2025, users found narrow no-break spaces (U+202F) in some ChatGPT o3 and o4-mini answers.
💡 Note
ChatGPT text can carry a watermark too. In October 2026, OpenAI began rolling out textGrain, which marks eligible ChatGPT and Codex output for users in the EU. Google also released SynthID Text to outside developers in October 2024, so it is no longer a Gemini-only system
Which AI Tools Watermark Text (October 2026)
As of October 2026, the answer to which AI chatbots watermark text covers three providers: Google, OpenAI and Anthropic. Microsoft Copilot and DeepSeek have no documented text watermark.
Provider | Text watermark? | Scope | Who can check it |
|---|---|---|---|
Google Gemini | ✅ Yes, SynthID Text | Text from the Gemini app and web experience; SynthID Text is also available to other developers | No public text checker; the public SynthID Detector checks images, video and audio only |
OpenAI ChatGPT / Codex | ✅ Yes, textGrain (announced October 5, 2026) | Eligible ChatGPT and Codex output for EU users, rolling out over several weeks; API opt-in worldwide; not a global default | Approved researchers and expert organizations only |
Anthropic Claude | ✅ Yes (models launched from August 2, 2026) | Supported models worldwide, across Claude, Claude Code, Claude Cowork, Claude Tag and the API; older models being added | Detection API in private preview; free Claude Content Checker for files (C2PA) |
Microsoft Copilot | ❌ No documented text watermark | A Claude model behind Copilot can carry Claude's mark | n/a |
DeepSeek | ❌ No documented statistical watermark | DeepSeek's visible AI label under China's labeling rules | n/a |
Provider details were checked on each company's own pages in November 2026. Rollouts change quickly, so confirm current scope on the provider's site.
How to Remove AI Markers and Hidden Characters in Microsoft Word
To remove AI markers from a Word document, paste the text into a new file as plain text, then use Find and Replace with Word's special codes. The codes delete invisible characters and swap odd spaces for normal ones. Clear leftover formatting last, then run Document Inspector to strip file metadata.
Paste as plain text into a new document. Use Paste > Keep Text Only, or press Ctrl+Shift+V (Cmd+Shift+V on Mac) in Microsoft 365. Stray fonts, colors and the old file's metadata stay behind, but invisible Unicode characters come along, so keep going.
Show hidden marks. Click the ¶ button on the Home tab (Ctrl+Shift+8 on Windows). Non-breaking spaces show as a small raised circle, and zero-width characters show as small empty boxes or not at all, which is why step 3 matters.
Find and replace each hidden character. Press Ctrl+H (Control+H on Mac), type a code from the table below into "Find what" and fill in "Replace with." Click Replace All, then repeat for each row.
Clear leftover formatting. Select all text, then click Home > Clear All Formatting (the "A" with an eraser).
Remove file metadata. On Windows, go to File > Info > Check for Issues > Inspect Document. Tick Document Properties and Personal Information, click Inspect, then click Remove All. On Mac, go to Tools > Protect Document and tick "Remove personal information from this file on save."
Read it back. Keep ¶ turned on and check that the raised circles and small boxes are gone.
Character | Unicode | Word "Find what" code | Replace with |
|---|---|---|---|
Zero-width space | U+200B | ^u8203 | Nothing |
Zero-width non-joiner | U+200C | ^u8204 | Nothing |
Zero-width joiner | U+200D | ^u8205 | Nothing (skip if the text has emoji, since some emoji use it) |
Word joiner | U+2060 | ^u8288 | Nothing |
Byte order mark | U+FEFF | ^u65279 | Nothing |
Non-breaking space | U+00A0 | ^s | One normal space |
Narrow no-break space | U+202F | ^u8239 | One normal space |
Soft (optional) hyphen | U+00AD | ^u173 | Nothing |
💡 Tip
Word's ^u code takes the character's decimal number, not its hex Unicode code. That's why U+200B becomes ^u8203 in the "Find what" box.
No Word setting removes a statistical watermark, because that signal is part of the word choice. Copilot drafts follow the same rule: Microsoft documents no text watermark of its own, but a Claude model behind Copilot can carry Claude's mark, as the guide to what Copilot does to text in Word explains.
How to Clean AI Markers and Hidden Characters in Google Docs
To clean AI markers in Google Docs, paste the text without formatting, open Find and replace, and tick "Use regular expressions." One pattern then deletes zero-width characters, and a second swaps odd spaces for normal ones. Clear the formatting last with Ctrl+\.
Paste without formatting. Press Ctrl+Shift+V (Cmd+Shift+V on Mac).
Open Find and replace. Go to Edit > Find and replace (Ctrl+H, or Cmd+Shift+H on Mac) and tick Use regular expressions.
Delete zero-width characters. Type [\u200B-\u200D\u2060\uFEFF\u00AD] into "Find," leave "Replace with" empty, and click Replace all.
Fix odd spaces. Type [\u00A0\u202F] into "Find," type one normal space into "Replace with," and click Replace all.
Clear formatting. Go to Format > Clear formatting, or press Ctrl+\ (Cmd+\ on Mac).
Check the download before sharing. A .docx you export from Google Docs carries its own file properties, so run Word's Document Inspector (step 5 in the Word steps) before you send it.
💡 Tip
Google Docs reads each \u code as one character, so \u200B means the zero-width space. If your text has emoji, remove \u200D from the first pattern, because some emoji need the zero-width joiner.
Prefer one step? Phrasly's AI Text Watermark Remover cleans hidden characters and stray formatting out of a pasted draft in one pass.
Why You Can't Find and Replace a Statistical AI Watermark
A statistical AI watermark is invisible and has no character to find, so find and replace can't touch it. SynthID Text, textGrain and Claude's watermark use a secret key to nudge which words the model picks, which spreads the signal across the whole passage. Plain-text pasting and metadata tools leave it in place too, and only new wording changes it.
SynthID-Text is not stored in metadata or embedded as its own code. It is distributed throughout the text.
This matters because the watermark is distributed as a statistical watermark pattern embedded across thousands of tiny decisions about which words were slightly more or less likely to be generated.
It's not stored in any single sentence. The watermark is in the document as a whole.
Rewriting one paragraph does not break that pattern. Neither does substituting some synonyms and switching around two sentences. The signal persists through small edits because it's systemic.
💡 Note
C2PA (Coalition for Content Provenance and Authenticity) is an open standard for signed provenance metadata on files such as images. Anthropic now attaches C2PA credentials to supported files Claude creates, such as PNG and JPEG images. These credentials belong to the file, so they don't come along when you copy the text.
💡 Note
In October 2026, Google opened its SynthID Detector portal to everyone. It checks images, video and audio, including media from partners such as OpenAI, NVIDIA and Kakao, but it doesn't check text.
💡 Note
Google's Nature paper on SynthID Text tested the watermark on nearly 20 million Gemini responses. User feedback on watermarked answers matched feedback on unwatermarked ones, which showed the watermark didn't lower text quality.
Does Paraphrasing Remove an AI Watermark?

Light paraphrasing usually doesn't remove an AI watermark, because a few synonym swaps or reordered sentences leave most of the statistical pattern in place. Google, OpenAI and Anthropic all say heavy rewriting and translation weaken detection. That's also why a missing watermark never proves a human wrote the text.
Paraphrasing simply replaces superficial words, but the watermark pattern is not embedded in the surface words themselves.
Rather, the watermark pattern lives in the statistical distribution of word choice throughout the passage. Replace 'Utilize' with 'use' and 'demonstrate' with 'show,' but that pattern remains unchanged.
SynthID Text is built to survive light edits, such as cropping pieces of text or changing a few words. Google says detector confidence "can be greatly reduced" when text is thoroughly rewritten or translated, according to Hugging Face's SynthID Text post. The model doesn't choose words randomly; it chooses them from shifted probability distributions that perpetuate the watermark.
The same principle holds true for simple sentence reordering. You aren't replacing the statistical material, you're reordering it.
💡 OpenAI's Own Test
In OpenAI's evaluation of 400-token passages, replacing 10% of words with synonyms cut textGrain detection from about 92% to 66%, according to OpenAI's textGrain announcement. Replacing 25% of words cut it to 17%.
There's a catch with paraphrasing tools: they run your text through another AI model. If that model watermarks its own output, the result can carry a new watermark. Anthropic says Claude output can carry a Claude mark even when the original text came from somewhere else, so this applies to text Claude only proofreads, translates or summarizes.
💡 Note
According to Google DeepMind's SynthID documentation, the watermarking system was designed to be robust to light, common edits.
What Doesn't Work: A Quick Summary
Paraphrasing: surface-word swaps leave the statistical distribution intact
Synonym substitution: a few swaps barely move the pattern
Sentence reordering: reshuffles the same material without replacing it
Minor manual editing: the watermark pattern is holistic, not sentence-level
Pasting into Notepad or as plain text: clears formatting, not invisible characters or statistical patterns
What Does a Better Approach Actually Look Like?
If you know why doing the obvious doesn't work, you know what will. It's not aggressive paraphrasing. It's operating on another level of the document.
Tier 1: Light Restructuring (Limited Effectiveness)
Moving words around, tweaking paragraphs, substituting a few words here and there. Honestly, that's low-impact.
The underlying statistical watermark pattern stays intact through surface-level changes. Do it if you're already editing heavily. It won't fix the problem by itself.
For hidden characters and formatting, the Word or Google Docs steps above do the job, while surface edits still do nothing for a statistical watermark.
Tier 2: Idea-Level Rewriting (More Effective, More Time-Intensive)
The big change is that you work from the concept, not the draft.
Instead of asking yourself, "How can I rewrite this sentence?" ask yourself, "What do I really want to say here and how would I say it in normal conversation?"
Rewrite the logic and flow of the piece, rather than simply the words.
This is what clean AI text looks like. You use the AI draft as a guide/comparison, not a copy/paste doc. It will take longer, but this is how you make the work sound like you wrote it.
If you've drafted something with Gemini, GPT or Claude and want to make it 100% yours, this is the baseline for rewriting.
Tier 3: Phrasly's AI Text Watermark Remover (A Clean Pass for Hidden Characters)
Phrasly's AI Text Watermark Remover cleans hidden characters, such as zero-width spaces, odd non-breaking spaces and similar AI markers, along with formatting artifacts from pasted text. It returns clean, natural text you can keep editing, and it handles long drafts.
It complements idea-level rewriting, not replaces it. Use it as the final technical pass after your own rewrite.
A Note on Ethical Use
This tool is intended to support honest, transparent workflows not to misrepresent the origin of content in contexts where disclosure is required. Responsible use includes situations such as:
Refining AI-assisted drafts you genuinely authored. For example, a freelance writer who used AI to brainstorm and outline, then wrote the final copy themselves, and wants the language to reflect their natural voice consistently throughout.
Publishing content on platforms with overly sensitive detectors such as a small business owner who wrote a blog post with light AI assistance but keeps getting flagged by blunt detection tools that cannot distinguish between AI-generated and AI-assisted human writing.
Preparing professional documents for personal use. For instance, a non-native English speaker who uses AI to help express their own ideas more clearly, and wants the final resume or cover letter to read as naturally as possible without surface-level detection artifacts.
Finished the rewrite in your own words? Run one last clean pass so no stray characters or hidden formatting stay in the final draft. 👇
Approach Comparison
Method | Hidden characters | File metadata | Statistical watermark | Effort |
|---|---|---|---|---|
Synonym swaps / reordering | ❌ No | ❌ No | ⚠️ Barely | Low |
Paste as plain text | ❌ No | ✅ Yes (new file) | ❌ No | Low |
Find and replace (Word / Google Docs) | ✅ Yes | ❌ No | ❌ No | Medium |
Word Document Inspector | ❌ No | ✅ Yes | ❌ No | Low |
Idea-level rewrite in your own words | ✅ Yes (new text) | ✅ Yes (new file) | ⚠️ Weakens it (per vendors) | High |
Phrasly's AI Text Watermark Remover | ✅ Yes | n/a (text in, text out) | ⚠️ Not promised | Low |
Is It Your Right to Edit the AI Text You Created?
Yes, with clear boundaries. You created the prompt and set conditions for the output, and Google's Terms of Service say Google "won't claim ownership" of original content you generate with its services.
That watermark just monitors where it was generated from. Provenance, not copyright.
Editing AI-generated text as part of your job is no different than editing a ghostwritten manuscript. The ethical dilemma of editing it doesn't stem from your ability to do so, but rather from whether that editing constitutes deception.
⚖️ The EU AI Act in 2026
The EU AI Act is already law, and its Article 50 transparency duties apply from August 2, 2026. AI providers must mark their output in a machine-readable way, which is why textGrain and Claude's watermark arrived in 2026.
Anyone who publishes AI-generated text to inform the public on matters of public interest must disclose it. The exception is text that a human reviewed and that a person or company holds editorial responsibility for. The Commission's Code of Practice on marking and labelling AI-generated content, published in June 2026, sets out practical steps for these duties. This isn't legal advice.
There's more nuance to unpack on the ethics side, especially around Gemini. Is It Ethical to Remove a Gemini Watermark? is worth a read if you want the complete picture.
Ethical Considerations: When Removing an AI Watermark Is and Isn't Okay?
The ethical concern is not the act of removing a watermark. It's the intent and context behind it. Here's a direct framework.
Where It's Completely Fine
✅ Editing AI-assisted drafts you created prompts for, before publishing on your own blog, site, or social media.
✅ Cleaning up AI-generated copy as part of a freelance or marketing workflow where AI use is permitted.
✅ Making AI-drafted text sound more like your voice before sending it to a client.
✅ Using a dedicated tool like Phrasly's AI watermark remover as part of a normal professional writing and editing process.
Where the Line Is
❌ Submitting AI-generated work where explicitly prohibited crosses an ethical line. The breach is the deception, not the watermark removal itself.
❌ Publishing AI content in high-stakes contexts (medical, legal, political) designed to appear human-authored creates real risks around responsible AI practices.
❌ Using removal to deceive a platform or reader with a legitimate need to know content is AI-generated.
❌ Removing marks to hide AI use from a school, client or employer that requires disclosure.
How to Remove AI Watermarks From Your Own Text: A 5-Step Workflow

You can get rid of AI watermarks in your own text in five steps: identify the type, clean hidden characters, strip file metadata, rewrite the ideas in your own words, and read the result back. The cleaning steps handle AI markers and metadata, and your rewrite is what makes the writing genuinely yours.
Identify what you're dealing with. Learn how to detect an AI watermark first, because a hidden character, a metadata field and a statistical pattern each need a different fix.
Clean hidden characters and formatting. Use the Word or Google Docs steps above, or paste the draft into Phrasly's AI Text Watermark Remover.
Strip file metadata. Paste the text into a fresh document, or run Word's Document Inspector.
Rewrite at the idea level in your own voice. Follow Tier 2 above and work from what you want to say, not from the AI's sentences.
Read it back, then run a free check. Ask yourself whether it sounds like you wrote it. A detector reads writing style, not watermarks, so use the check below to find generic passages worth another rewrite.
If you're using AI as a drafting tool and editing its output as part of a typical professional writing workflow, then restructuring it is valid
Frequently Asked Questions
Does Paraphrasing Actually Remove an AI Watermark From Text?
Light paraphrasing rarely does. Paraphrasing changes the words on the surface without altering the underlying statistical watermark. Google, OpenAI and Anthropic say heavy rewriting or translation weakens detection, so rewrite at the idea level instead.
What Is an Invisible AI Watermark, and Why Can't I Remove It Manually?
An invisible AI watermark comes in two kinds. Invisible characters, such as zero-width spaces, can be removed by hand with the Word or Google Docs steps above. A statistical pattern like SynthID Text has nothing to delete, because the signal is spread across the model's word choices.
Is It Legal or Ethical to Edit AI Watermarks Out of Text I Wrote?
Yes, when you edit your own AI-assisted writing in settings where AI use is allowed. The line is crossed when the intent is to deceive. Under EU AI Act Article 50, some public-interest text must be disclosed as AI-generated from August 2, 2026. This isn't legal advice.
What's the Difference Between Removing an AI Watermark and Bypassing AI Detection?
Removing a watermark means refining AI-assisted writing you produced as part of a standard editing workflow. "Bypassing AI detection" suggests trying to deceive tools like GPTZero or Turnitin. Those detectors score writing style, and they don't read provider watermarks.
Can I Remove an AI Watermark From Text for Free?
The manual Word and Google Docs steps cost nothing. A free Phrasly account cleans a first short draft, and longer jobs need the trial or the Unlimited plan on Phrasly's pricing page. The tool doesn't promise to remove provider watermarks like SynthID, textGrain or Claude's, which are part of the wording.
Does 'Remove GPT Watermark' Mean ChatGPT Actually Embeds One?
For some ChatGPT text, it now does. Since October 5, 2026, OpenAI has been rolling out its textGrain watermark to eligible ChatGPT and Codex output for EU users, though it isn't a global default. Claude has marked text since August 2026, as explained in how Claude's text watermark works.
Can File Metadata Show That I Used AI, and How Is It Different From Hidden Characters?
File metadata can show AI use on some files. Metadata is stored with the file: author, edit time, app name and, on some AI-made files, signed C2PA credentials. Hidden characters are part of the text and travel when you copy it, while metadata stays with the old file.
Can You Remove SynthID From Text?
No tool can honestly promise to remove SynthID from text, because there's no character or tag to delete. Google says heavy rewriting or translation weakens detection. The honest route is to rewrite in your own words and disclose AI use where it's required.

Written by
Muhammad Usman Ali
Pakistan
Muhammad Usman Ali is an experienced SEO content writer with 3+ years of professional writing experience. He specializes in AI tools, AI detection technologies, and search engine optimized content.


