Does ChatGPT Watermark Text? AI Watermarks Explained

Short answer: yes, in some cases, and it changed this year. Since early October 2026, ChatGPT text produced for users in the EU is being watermarked. Anthropic said in August that Claude's text output carries a watermark worldwide. And Google has been watermarking Gemini app text with SynthID since 2024.

A lot of what's circulating about this is wrong. Here's what each company has actually confirmed, who can check for the marks, and what they can't tell anyone.

Does ChatGPT watermark text?

Yes, for some users. On October 5, 2026, OpenAI said it would add an invisible watermark to text generated by ChatGPT and Codex in the European Union, rolling out over the following weeks to eligible users on all plans, according to TechCrunch. The stated reason is compliance with the EU AI Act.

A few details matter here:

• It's EU only for ChatGPT, for now. OpenAI said it isn't making text watermarking a global default at launch.
• The API is opt-in. Developers anywhere can switch it on for select models, but it's off by default.
• It doesn't identify you. OpenAI says the watermark carries no information about the user.
• It's called textGrain. OpenAI published a technical report on the method, written with researchers from the University of Pennsylvania and Yale. A secret key nudges which word the model picks next, and a detector holding that key looks for the resulting pattern.

Outside the EU, a regular ChatGPT conversation isn't covered by this rollout as of late October 2026. That could change, so check OpenAI's own announcements.

Does Claude watermark text?

Yes, and more broadly than ChatGPT. On August 11, 2026, TechCrunch reported that Anthropic had updated a support page to say it would watermark text from its models. Anthropic followed with a longer explainer, How Claude's text watermarking works, dated August 14.

Here's what Anthropic says in that post and on its support page:

• It's global. Anthropic applies it worldwide because it doesn't yet have a durable way to limit it by region.
• It's applied at the model level, so it shows up across the Claude apps, the API, Claude Code and its other products, including Claude offered through cloud partners.
• It's based on Google DeepMind's SynthID-Text approach.
• Older models are being phased in. Models launched on or after August 2, 2026 support marking at launch. Earlier models are getting it during a legal transition period.
• Files are different. Images and other generated files get signed C2PA Content Credentials in their metadata. That's a label attached to the file, not a watermark in the words.

What about Gemini and SynthID text?

Google got here first. In May 2024, Google DeepMind said it was extending SynthID to text in the Gemini app and web experience. In October 2024 it published the method in Nature and made the text watermark open source. MIT Technology Review reported that a live test across about 20 million Gemini responses found users didn't notice a drop in quality.

The reference code is on GitHub, and Google's developer docs say a production version ships in Hugging Face Transformers from v4.46.0. Open source here means any developer can add SynthID watermarks to their own model. It doesn't mean anyone can check Gemini's text. The docs tell developers to keep their watermark configuration private, and Google holds its own.

Statistical watermarks vs hidden characters

Statistical watermarks (what OpenAI, Anthropic and Google use)

A language model writes by picking one token (a word or piece of a word) at a time from a list of likely options. Often several options are about equally good. A statistical watermark uses a secret key to tilt those close calls in a way that looks random to a reader but forms a pattern a detector can measure across a whole passage.

Nothing is added to the text. Anthropic says plainly that there are no hidden characters. The watermark is in which words were chosen, so it travels with the text when you copy and paste it.

Hidden characters (a different thing entirely)

Hidden characters are invisible Unicode symbols, like zero-width spaces or narrow no-break spaces, that end up in copied text. In April 2025, the education startup Rumi reported narrow no-break spaces (U+202F) in longer answers from OpenAI's o3 and o4-mini models. OpenAI told Rumi they weren't a watermark, calling them a quirk of large-scale reinforcement learning. A day after posting that reply, Rumi retested and found them gone.

Stray characters like these can cause real formatting trouble: odd line breaks, broken find and replace, word counts that don't add up. But they aren't how any of the 2026 watermarks work.

Character cleaners don't remove watermarks

To be clear: a character cleaner does not remove a statistical watermark. Stripping invisible symbols, curly quotes or Markdown leaves every word exactly where it was, and the words are the watermark.

Our free AI text cleaner is a formatting tool. It tidies up invisible Unicode characters, em dashes, stray Markdown and spacing so pasted text behaves in Google Docs, Word or a CMS. It doesn't touch, detect or affect watermarks. Anyone claiming a one-click cleaner "removes the ChatGPT watermark" is describing something that doesn't exist.

Who can actually detect these watermarks?

Not you, and not most AI checkers. Detection needs the provider's secret key, and access is tightly held for now:

Provider	Whose text is watermarked	Who can detect it
OpenAI (textGrain)	ChatGPT and Codex users in the EU; API customers who opt in	Approved researchers and expert organizations, by application
Anthropic (Claude)	Supported Claude models, worldwide	Private preview for eligible organizations such as regulators, media, fact-checkers, researchers and educational organizations
Google (SynthID)	Gemini app and web text	Google. Its SynthID Detector portal, announced in May 2025, started with early testers and a waitlist for journalists, media professionals and researchers

Anthropic's support page says outright that there's no public watermark detection for text. And each company's detector only finds its own mark. A Claude watermark detector isn't built to spot a ChatGPT watermark, and the reverse is true too.

That has a practical consequence if you're worried about school or work. A typical AI detector, including our free AI detector, doesn't read watermarks at all. It estimates likelihood from writing patterns like predictability and sentence rhythm. For the background on how that works and why it misfires, see how AI detection works.

What watermarks can't tell anyone

All three companies are unusually frank about the limits, and the limits matter more than the headlines.

• Short text is hard. Anthropic says detection is unreliable on short passages. Google says SynthID is less effective on factual answers, where there's little room to vary wording.
• Some content barely carries a mark. Anthropic says watermarks are sparse on factual text, light proofreading and code that has to be exact. OpenAI says detection is weaker in areas like maths, where word choice is constrained, per 9to5Mac's summary of its report.
• Edited and translated text is harder to verify. In OpenAI's own tests on 400-token passages, detection fell from about 92% to 66% when 10% of the words were swapped for synonyms. Google says confidence drops a lot after thorough rewriting or translation.
• A missing watermark proves nothing. OpenAI and Anthropic both say the absence of a mark doesn't show a human wrote the text.
• A found watermark doesn't prove authorship either. Anthropic says a detected mark only means Claude was likely involved at some point. It can't tell whether Claude drafted the piece, translated it or just fixed the grammar.

That matters. If you ask Claude to proofread an essay you wrote yourself, parts of the result may carry a watermark. A watermark is evidence of involvement, not of how much of the thinking was yours. If your school or employer has an AI policy, follow it and be open about how you used AI tools. That protects you far better than guessing what a detector might say.

Why this is happening now: the EU AI Act

Article 50(2) of the EU AI Act requires providers of AI systems that generate synthetic text, images, audio or video to mark outputs in a machine-readable format so they're detectable as AI-generated. It applies from August 2, 2026. There's an exception for systems doing an assistive function for standard editing, or that don't substantially alter the input.

The Digital Omnibus on AI (Regulation (EU) 2026/1744) gave systems already on the market before August 2, 2026 until December 2, 2026 to comply. That explains the staggered rollouts: new models are marked at launch, and older ones are being brought in before the December deadline.

What this means if you write with AI

Day to day, not much changes. Watermarks don't alter how your text looks, and most readers can't check for them. But the direction is clear: AI providers now have a legal duty to make their output detectable, and detection access is opening up to regulators, researchers and educational organizations.

The sensible response isn't hunting for workarounds. It's using AI as a drafting aid, putting your own thinking and voice into the final version, and being honest about your process where it matters. That's the idea behind WriteHumanly: helping AI-assisted drafts read naturally, in your voice, with every fact kept intact.

Frequently asked questions

Does ChatGPT watermark text outside the EU?

Not by default as of late October 2026. The ChatGPT and Codex rollout is EU only. API developers anywhere can opt in on select models, so text from apps built on OpenAI's API may be watermarked if the developer has switched it on.

Does Claude watermark text?

Yes. Anthropic says supported Claude models watermark text worldwide, across its apps, API and Claude Code. Models launched on or after August 2, 2026 are marked at launch, and older models are being added during the EU transition period.

Can I check whether my text has a SynthID or ChatGPT watermark?

Generally, no. Detection needs each provider's secret key, and access is limited to approved researchers, regulators and similar organizations. A free online "watermark checker" can't read these statistical marks. At most it finds invisible characters, which are a separate thing.

Does removing hidden characters remove the watermark?

No. The watermarks from OpenAI, Anthropic and Google live in word choice, not in hidden symbols. Cleaning out invisible characters fixes formatting problems but leaves any watermark exactly as it was.

Will Turnitin or GPTZero detect AI watermarks?

None of the providers' announcements name Turnitin or GPTZero as having detector access, and a watermark can only be checked with the provider's own key. Access is changing quickly, though, so check each provider's current documentation rather than assuming either way.