AI Text Watermark Removal

People searching for AI text watermark removal are usually chasing one of three unrelated things. Conflating them is where most of the confusion starts.

• Cleaning invisible Unicode out of pasted text. This is formatting hygiene, not defeating anything a provider built on purpose, and you can verify it yourself in seconds.
• Reducing a statistical watermark's detector confidence by rewriting, paraphrasing, or translating. The research literature treats this as a live arms race, not a problem solved in either direction.
• Stripping file-level C2PA provenance metadata. Closer to routine file handling than to breaking a cryptographic scheme, since ordinary uploads and screenshots already destroy it.

Only the first has a reliable, checkable answer. Everything after that depends on which provider wrote the text and how much of it you have.

Anthropic says compatible Claude models weave an imperceptible watermark directly into generated text. It applies wherever Claude is offered, worldwide, because, per Anthropic's help centre, watermarking is applied at the model level and is present no matter which Claude product or surface the text comes from.

• The Claude app, Claude Code, Claude Cowork, and Claude Tag
• The Claude API
• Claude models accessed through AWS, Google Cloud, and Microsoft Foundry

Models launched on or after August 2, 2026 support marking at launch. Older models sit in a transition period, and Anthropic's help page now dates its end: every model released before August 2, 2026 is to be covered by December 2, 2026. Anthropic's 2026-09-25 email to Claude Platform customers says Claude Opus 5 carries the mark from September 14 (its 2026-09-04 notice to administrators had said September 9) and schedules Fable 5, Sonnet 5 and Opus 4.8 from 2026-09-30.

No detector the public can run exists: Anthropic's shipped on 1 September 2026 into a private preview for eligible organizations under EU law. Even for those who get access, Anthropic is explicit that a match would only mean text may have been processed by Claude, not proof that it was.

Google documents SynthID Text as a logits processor applied during generation after standard sampling, and open-sources a production-grade version with a reference detector. Google's own page says the Gemini app and web experience carry the watermark.

Then a Google-affiliated reply on the company's developer forum, dated August 5, 2026, said Gemini API text carries no watermark at all and that native text watermarking is not planned. On August 19 the same account reversed it, saying API text "IS actually SynthID-watermarked" and that its earlier statement "was incorrect." No Google documentation supports either version for the API, so this is still unresolved, just in the opposite direction.

OpenAI is the clean case. It has confirmed C2PA and SynthID for images, SynthID alone for audio, and nothing for ordinary ChatGPT text, and has published unusually specific reasoning for holding back a text watermark it already built.

The most-cited case against watermarking is Sadasivan et al.'s 2023 paper, now past 600 citations, built on a recursive paraphrasing attack. Its conclusion is the one most people already assume: a motivated paraphraser beats these detectors.

Then the original scheme's own authors stress-tested it. Kirchenbauer et al.'s 2024 follow-up found the mark stayed detectable after strong paraphrasing once roughly 800 tokens were available, at a false-positive rate of one in 100,000.

So whether paraphrasing wins depends on the scheme and on how much text there is. It is not a clean yes or no.

SynthID Text has drawn three independent evaluations since Google shipped it, and none of them is flattering.

• An EMNLP 2025 paper found its resilience to paraphrasing and translation similar to a much older, simpler baseline, with no clear advantage.
• A 2025 IEEE TrustCom paper called it susceptible to paraphrasing, copy-paste edits, and back-translation, and built a hybrid defense that improved detection by 11.1% on average.
• A 2026 preprint found that 98.3% of initially-detected watermarked texts lost their signal after paraphrasing.

One caveat applies to all three. They test the open-sourced algorithm or a reimplementation, never Google's production configuration, which Google has never published, so read them as limits of the published method rather than measurements of Gemini.

The ETH Zurich result cuts deeper than paraphrasing does. Approximating a provider's secret green and red token rule from public API queries for under $50 pushed scrubbing success from roughly 0% to better than 85% in settings previously assumed safe.

A follow-up black-box method from the same lab matched that without even knowing which scheme was running. On defense, a 2025 paper from that lab found current spoofing attacks leave detectable artifacts of their own, so the arms race runs in both directions.

A hobbyist blog post from April 2026 claims FFT-based spectral analysis reverse-engineers Google's image SynthID watermark well enough to detect it 90% of the time and remove it 91% of the time while preserving image quality. The same post claims over 90% success scrubbing text watermarks through paraphrasing, with code published on GitHub.

None of it has been peer-reviewed, replicated, or corroborated. When the project hit Hacker News, commenters noted it had only been tested against the project's own detector rather than Google's.

A follow-up thread quoted the researcher's own admission that an earlier version reached a 16% evasion rate, which is a long way from cracked. Worth knowing the claim exists, since it surfaces in searches and forum threads. Not worth building a strategy on.

A Reddit thread asking plainly how to remove Claude's text watermark split into two camps. One treated the question itself as suspicious; the other argued that paying for output should mean the right to edit it freely.

Two removal methods came up in the replies:

1. Rewriting the text by hand.
2. Running it through a second model in the hope of adding noise.

Neither is documented. Both are commenter speculation, and heavy rewriting risks destroying the writing along with whatever mark it carried.

What that conversation misses is that Article 50 never required marking on everything. The exemptions are broad:

• Short sequences of numbers, symbols, or letters
• Source code
• Machine-to-machine output a human never sees
• Closed-loop industrial processes, short of their final output
• Standard editing-assistive functions

A lot of everyday AI text was never going to carry a mark in the first place.

1. Scan for hidden Unicode first. It's the one layer you can verify yourself, and the only one where removal means something definite.
2. Identify the source before drawing conclusions. Claude, Gemini's app and web experience, and ChatGPT sit at genuinely different confirmed statuses, and Gemini's API contradicts Google's own marketing.
3. Weigh the length. The strongest published survival result needed roughly 800 tokens, so a short passage was weakly marked before anyone touched it.
4. Hold every removal claim to one standard: tested against a real detector and corroborated by someone else, or it's just a claim on a page.