ChatGPT · Claude · Gemini · updated 20 October 2026
Remove AI text watermarks
Paraphrasing erased the mark from 98.3% of the texts where it had first been found, in a July 2026 preprint that tested an open implementation of SynthID-Text. The makers put it in their own words: Anthropic says a complete rewrite that replaces every word removes the Claude mark, and OpenAI says editing can weaken its signal; only the makers hold the keys that test those marks. Paste your text below: the free cleaner clears invisible characters, and the rewrite tab gives the passage new wording.
Short answer
AI text watermarks sit in word choice, not in extra characters, so deleting characters leaves them where they are. Anthropic says a complete rewrite replacing every word removes the Claude mark, and OpenAI says editing can weaken its signal. Only OpenAI, Anthropic and Google hold the keys that test their marks, so no public tool shows a mark's status. The free cleaner above clears invisible characters, and the rewrite tab returns the text in new wording.
Which AI makers watermark their text today?
Three of them. OpenAI is adding a mark called textGrain to eligible ChatGPT and Codex text in the EU, Anthropic marks Claude text worldwide, and Google has marked Gemini app and web text with SynthID-Text since May 2024.
| Maker and product | Method, as the maker names it | Where it applies, as announced | Announced |
|---|---|---|---|
| OpenAI: ChatGPT, Codex | textGrain, "an invisible statistical signal" in the model's word choices | Eligible EU text, added "over the coming weeks"; API customers anywhere can opt in for select models, off by default | 5 Oct 2026 |
| Anthropic: Claude | A version of Google DeepMind's SynthID-Text approach | Applied globally at launch, because Anthropic says it cannot yet scope the mark by region | 14 Aug 2026 |
| Google: Gemini app and web | SynthID-Text, which shifts word probabilities while the answer is generated | The Gemini app and web experience; this page does not stretch that to other Google products | 14 May 2024 |
OpenAI leaves "eligible" undefined, and no report we found says the EU rollout has finished. The three designs share one trait: a secret key held by the maker shapes the wording at the moment of writing. Anthropic says plainly that nothing is added to Claude text and there are no hidden characters. Google never words it that way, but the Nature paper behind SynthID-Text says the method "modifies only the sampling procedure", so, in our reading, there is no character to find.
Sources: OpenAI, 5 Oct 2026 · Anthropic, 14 Aug 2026 · Google DeepMind, 14 May 2024 · Dathathri et al., Nature, 2024 (all read 19 Oct 2026)
What does each maker say removes its watermark?
None of the three says a few edits are enough. OpenAI says editing can weaken its signal, Anthropic says a complete rewrite that replaces every word removes the Claude mark, and Google says its detector's confidence can be greatly reduced when text is thoroughly rewritten or translated.
| Maker | Light editing | Full rewrite | Translation | Where the mark is thinner |
|---|---|---|---|---|
| OpenAI | "Editing can weaken the watermark." Its partial-swap test on 400-token passages: synonyms for a quarter of the words took detection from about 92% down to 17% | No rule stated | Not stated | Short passages, and content such as mathematics |
| Anthropic | "probably won't remove the watermark completely" | Removed, it says, by "a complete rewrite where every word is replaced" | A translation Claude itself writes carries the mark, since Claude picks every word | Small samples, factual passages, code |
| Holds up under "modifying a few words and mild paraphrasing" | Detector confidence "can be greatly reduced" after a thorough rewrite | The same sentence names translation into another language | Factual responses |
Read side by side, the rows point one way: each maker ties a weaker or missing mark to how much of the wording changed, and OpenAI reports detection already falling at a 10% synonym swap and further at 25%. The two translation cells describe different cases (Claude translating, versus text later translated), so they do not contradict each other.
For how the Claude mark works, with a live meter that shows how much of your wording survived, see the Claude watermark remover.
Sources: OpenAI, 5 Oct 2026 · Anthropic, Claude text watermark · Google AI for Developers, SynthID · Google DeepMind, 14 May 2024
Which fix does my text need: cleaning or rewriting?
Cleaning deletes characters you cannot see; rewriting changes the words themselves, the layer where the makers say their marks sit.
| What you see or want | What fixes it |
|---|---|
| Odd gaps, a form or code editor that chokes on pasted text, a scan listing strange characters | The free tab. It catches 19 groups of invisible characters and 6 groups of unusual spaces, deletes the first and turns the second into a plain space, and runs in your browser with nothing sent |
| The wording itself should change, which is where all three makers put their marks | The rewrite tab. Sign in with an email code, paste at least 20 words, and the text is cleaned on the way in and again on the way out |
| Content Credentials attached to an image, audio clip or file | Not a text matter, and this tool leaves files alone. Claude's Content Checker reads files, not text |
One point of logic, ours rather than a vendor's: every maker's checker reads the sequence of words, so a tool that hands back the same words hands back the same pattern. Start with the ChatGPT watermark remover on the home page if your text came from ChatGPT; the same two tabs work for any AI text.
Sources: Claude help center, how Claude marks content · Nature, 2024
Does independent research agree that rewriting works?
Largely yes. Every study below, though, tested open versions of the methods, not the live systems at OpenAI, Anthropic or Google.
| Study | Date | What was tested | Result, as the authors report it | Not covered |
|---|---|---|---|---|
| Kirchenbauer et al., arXiv 2306.04634 | 7 Jun 2023 | The green-list watermark | After strong human paraphrasing, still detectable after about 800 tokens on average, at a 1e-5 false-positive rate | Any deployed mark |
| arXiv 2607.16010 (preprint) | 17 Jul 2026 | KGW, Unigram and the MarkLLM implementation of SynthID-Text; 846 valid paraphrase runs | 98.3% of SynthID texts first detected lost the mark after paraphrasing; 80% of untouched SynthID texts were missed from the start; 5.4% of paraphrased human texts were flagged | Gemini, Claude and ChatGPT in production |
| arXiv 2508.20228 (preprint) | 27 Aug 2025, revised 21 Oct 2025 | SynthID-Text against paraphrasing, copy-paste edits and back-translation | Such attacks "can significantly degrade watermark detectability" | No deployed system named |
| arXiv 2609.09604 (preprint) | 9 Sep 2026 | Open SynthID-Text on two open-weight models | The authors write that "no public tool can test the deployed systems" | Removal rates (it measures quality) |
The 2023 study is the counterweight: even after strong human paraphrasing, the older green-list mark could still be found once about 800 tokens of text were read, on average. The three newer papers are preprints and have not been peer reviewed.
Sources: arXiv 2306.04634 · arXiv 2607.16010 · arXiv 2508.20228 · arXiv 2609.09604 (abstracts read 19 Oct 2026)
How can you tell whether a watermark is gone?
Only the makers can say, because only they hold the keys that read their marks, and none of them runs a public checker for text.
OpenAI gives text checks to approved organizations, Anthropic's Claude detection sits in private preview for eligible organizations, and Google's public SynthID Detector reads images, video and audio, not text. Anthropic adds that a missing mark does not show the content was not AI-generated. What you can see yourself is what the tool did: the count of invisible characters the cleaner removed, and the new wording the rewrite tab returned.
Who holds each key, and what a positive or negative result proves, is set out on the watermark detector page.
Sources: OpenAI Content Provenance API · Claude help center · SynthID Detector
Is it allowed to remove an AI text watermark?
As we read Article 50 of the EU AI Act, nothing in it forbids editing AI text: its marking duty in 50(2) falls on the providers of AI systems. Disclosure duties under 50(4), and a school's or employer's own rules, can still apply to you.
Providers must keep outputs "detectable as artificially generated or manipulated". Under Article 50(4), deployers who publish AI text to inform the public on matters of public interest must say so, unless the text went through human review or editorial control and a person or company holds editorial responsibility for it.
Our advice, not a sourced rule: a school, employer, journal or client may set its own policy on AI help. Follow it, and where it asks you to disclose, disclose. This section is general information and should not be read as legal advice.
Source: European Commission AI Act Service Desk, Article 50 (read 19 Oct 2026)
What does the AI text watermark remover cost?
The invisible-character cleaner costs nothing and needs no account; rewrites begin with four free runs after an email sign-in, and paid plans start at $12 a month when billed yearly.
The free account covers four rewrites of up to 250 words each. Basic is $18 a month ($144 a year) with up to 1,000 words per rewrite, Pro is $27 ($216 a year) with 1,200, and Ultra is $48 ($432 a year) with 3,000. Payment runs through Stripe.
The rewrite tab hands back your passage in new wording and strips invisible characters from both your input and its output, so what you paste in comes back clean and freshly worded. Pick from the plans when the free runs are used up.