Does ChatGPT Watermark Text? Claude and Gemini Too (2026)

For two years the honest answer was “not really, except at Google”. That changed when the transparency rules of the EU AI Act started applying on 2 August 2026. Anthropic and OpenAI now mark their text invisibly too, or are about to. Here is what is actually live, what is rolling out, and what is rumor.

How a text watermark works

A language model writes one token at a time. At every step it has several plausible continuations: “fast”, “quick”, “in no time”. A statistical watermark uses a secret key to mark some of those options as “favored” at each position. The model leans slightly toward them, never far enough to pick a word that reads wrong.

In one sentence, that bias is invisible, even to a careful reader. Across a few hundred words it becomes measurable: human text contains roughly as many favored words as chance predicts, watermarked text contains noticeably more. Only whoever holds the key can run that count.

Two practical consequences:

• There is no hidden character to delete. The mark is the word choice.
• The signal weakens when the text is short, highly factual (a sum or a date leaves little choice), translated, or heavily reworded.

What is officially deployed

Google: SynthID Text. The first watermark deployed at scale. Google DeepMind has used it in Gemini since 2024 and published the method in Nature in October 2024, after a live test on around 20 million responses with no measurable quality loss. The code is open source, but Gemini’s production keys are not public, so nobody outside Google can check whether a passage came from Gemini. The public SynthID Detector, opened on 7 October 2026, checks images, audio and video, not text.

Anthropic: Claude. According to Anthropic’s help page, Claude embeds an imperceptible watermark in text from its supported models (the list is published), across its products and “wherever Claude is offered, worldwide”. Models launched on or after 2 August 2026 support it from day one. The mark travels with copy and paste and “may persist through some editing”, but Anthropic says it becomes unreliable when the passage is very short, or heavily edited, paraphrased, translated or mixed into other writing. The detector is in private preview for eligible organizations.

What is announced and rolling out

OpenAI: ChatGPT and Codex. OpenAI built a text watermark back in 2024 and chose not to ship it. In an August 2024 update, it said the method held up poorly against rewording by another model or translation, and could stigmatize some users, such as non-native English speakers.

On 5 October 2026, OpenAI announced textGrain: an invisible watermark added “over the coming weeks” to ChatGPT and Codex text, in the European Union only. API developers anywhere can switch it on, but it is off by default. Outside the EU, no default watermark on ChatGPT has been announced.

OpenAI publishes its own limits: replacing 10% of words with synonyms dropped detection from about 92% to 66%, and replacing 25% dropped it to 17%. The detector is limited to approved researchers and expert organizations.

The rumor: invisible spaces

In April 2025, the startup Rumi spotted narrow no-break spaces (Unicode U+202F) in long answers from o3 and o4-mini. Many read it as a hidden watermark. OpenAI said it was not one, but “a quirk of large-scale reinforcement learning”, and Rumi saw the characters disappear by 23 April.

OpenAI now states that textGrain adds no hidden characters, invisible spaces or unusual punctuation. Tools that “clean” those characters do not touch the real watermark, which lives in the words themselves.

Classic AI detectors are a different problem

Most online detectors have no access to any watermark. They guess, by measuring how predictable a text is. The output is a probability, not proof, and errors are common:

• OpenAI withdrew its own classifier on 20 July 2023 for its “low rate of accuracy”: it caught 26% of AI-written text and wrongly flagged 9% of human text.
• A Stanford study in Patterns (July 2023) found seven detectors classed, on average, 61% of TOEFL essays by non-native English writers as AI-generated.

Plain, regular or heavily formatted writing can be flagged while being entirely human.

Polish your text before you post it

“Clean your text” is a new feature in the CleanAi iOS app. It does a light rewrite: smoother phrasing, more varied sentence rhythm, fewer mechanical turns of phrase, same meaning. Use it on a caption, a post or a product description, whether you wrote it yourself or drafted it with an AI.

What it does, and what it doesn’t. A rewrite changes word choices, which can soften some of the signals statistical detectors look for and sometimes avoid a false positive on text you stand behind. It does not promise to make text “undetectable”, and it does not promise to remove a watermark. It is not built for homework, exams or academic work either: there, your institution’s rules apply. Get CleanAi.

Key takeaways

1. Gemini text has been watermarked since 2024, Claude text since August 2026, and ChatGPT text is being watermarked in the EU.
2. A text watermark is a slight statistical bias in word choice, not a hidden character.
3. The detectors for these watermarks are not open to the public, and public AI detectors are often wrong.

ChatGPT text watermark FAQ

Does removing invisible spaces remove the ChatGPT watermark?

No. According to OpenAI, the U+202F spaces spotted in 2025 were not a watermark, and textGrain adds no hidden characters. The mark is in the word choice itself.

Can an AI detector prove I used AI?

No. A statistical detector outputs a probability, with documented false positives. A watermark indicates that text from a given model is probably present, not who wrote it or how much human work went into it.

Does translating or rewording text remove the watermark?

Not necessarily, but it weakens it. Google, OpenAI and Anthropic all say translation, heavy paraphrasing and short passages make detection less reliable.

Related reading: what C2PA Content Credentials are · TikTok’s invisible watermark, explained