Does ChatGPT watermark text? What ChatGPT, Claude and Gemini do

Yes, but only in the EU. On 5 October 2026 OpenAI announced a ChatGPT watermark for text, rolling out to ChatGPT and Codex users in the EU over the following weeks; everywhere else it stays off unless an API customer turns it on. Anthropic watermarks Claude's text worldwide, starting with models launched since 2 August 2026, and Google has marked text in the Gemini app with SynthID since May 2024. None of these watermarks is made of hidden characters, and the public can't check text for any of them.

ChatGPT, Claude and Gemini text watermarks at a glance

As of 1 November 2026, from each company's own pages (listed under Sources).

How does a text watermark work?

A text watermark is a statistical pattern in the words a model picks. At many points in a sentence, several next words would be about equally good. A watermarking model lets a secret key steer that choice, so the finished text leans towards the key's pattern without reading any differently.

Anthropic says Claude's watermark is a version of SynthID-Text, the method Google DeepMind published in 2024, and that it only changes how Claude picks among words of similar quality. OpenAI's textGrain works on the same principle: its technical report ties each word choice to keyed randomness and notes that "the detector requires only the generated text and the secret key". Without the key there is nothing to see. With it, a detector counts how often the words line up with the pattern, which is why longer texts give clearer answers than short ones.

Does ChatGPT hide characters in its text?

No. The rumour that chatbots slip invisible characters into their answers as a watermark is older than any of these schemes, and it doesn't match how the companies describe them. Anthropic is explicit: "Nothing is added to the text and there are no hidden characters." OpenAI describes textGrain as an invisible statistical signal in the model's choice of words.

Zero-width spaces, odd spaces and look-alike letters do turn up in copied text, from PDFs, web pages and chat apps alike. They tell you how a text travelled, not who wrote it, and our free invisible character remover finds and cleans them in your browser. It also explains why an "AI watermark remover for text" that strips characters does nothing to these watermarks: the signal is in the choice of words, so there is no character to delete. What research and the law say about removing watermarks is in can AI watermarks be removed?

Can you check text for a ChatGPT or Claude watermark?

Not as a member of the public, as of 1 November 2026.

• OpenAI lets approved researchers and expert organisations apply for its text watermark detector, and grants access case by case. Given the risk of missed watermarks and false positives, it says it won't make the tool public at launch. Its public verify tool checks images and audio only.
• Anthropic runs its detection API as a private preview for eligible organisations: regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations, EU civil society groups, and businesses that must verify watermarks themselves. Its free Claude Content Checker reads Content Credentials in image, video and audio files, and Anthropic states plainly that "the tool does not check text."
• Google opened its SynthID Detector to everyone on 7 October 2026, for images, video and audio. It released SynthID Text as open source, so developers can watermark text from their own models.

Keeping text detection private is allowed under the EU's Code of Practice on transparency of AI-generated content, published on 10 June 2026. Because detection for text watermarks is less reliable, signatories may limit it, for a limited time, to verified expert users with a legitimate need, such as regulators, media, fact-checkers and researchers. Our guide to AI watermark detectors covers what you can check today and where.

What a missing watermark tells you

Very little. Anthropic puts it this way: "Lack of a detected mark doesn't mean the content wasn't AI-generated or processed." OpenAI says the same about ChatGPT text, and both list the usual reasons a mark goes undetected:

• the passage is too short, or mostly factual, so there were few word choices to steer;
• it was edited or translated after it was generated;
• it came from an older model, a model without watermarking, or another company's AI.

A detected watermark has limits too. It shows that the company's model generated or processed some of the text. It doesn't measure how much a person added, decide who owns the text or identify the user, and Anthropic notes that text can carry a Claude mark even when the ideas or wording came from somewhere else.

How well do text watermarks survive editing?

OpenAI is the only one of the three to publish figures for its text watermark. With its detector set to a 1% false positive rate, it found the watermark in about 95% of 400-token passages and about 80% of 200-token passages on topics like psychology, and far fewer on maths, where word choice is tighter. Editing weakens the signal quickly: replacing 10% of the words in a 400-token passage with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%.

Anthropic expects light editing to leave Claude's watermark in place, while a complete rewrite that replaces every word removes it. Google says SynthID Text's confidence scores can drop sharply when a text is thoroughly rewritten or translated, and that the method works less well on answers to factual questions.

Why the EU, and what about Switzerland?

The EU AI Act requires providers of AI systems that generate text, images, audio or video to mark the output in a machine-readable format, so that it can be detected as AI-generated (Article 50(2)). That duty has applied since 2 August 2026, and systems already on the market before then have until 2 December 2026 under the Digital Omnibus amendment of July 2026. The Commission's Code of Practice, signed by about 190 organisations including Anthropic, OpenAI, Google, Meta, Microsoft and Mistral, spells out how: free-form text longer than 200 tokens should carry an imperceptible watermark.

The companies drew the map differently. Anthropic applies its watermark everywhere, because it says it can't yet limit it by region reliably. OpenAI chose the EU only and says it is "not making text watermarking a global default at launch". Switzerland isn't in the EU, so OpenAI's rollout as announced doesn't reach ChatGPT users in Switzerland, or in the UK or the US, while Claude's watermark applies there as everywhere else.

Watermark checks versus AI detectors

A watermark check asks one narrow question: did this company's model, holding this key, produce some of these words? An AI detector asks what the writing itself looks like. IsThisAiWritten is that second kind: a classifier, not a watermark reader. It reads whatever text you paste, from any model or none, watermarked or not, and needs no key.

It is free, with no sign-up and no word limit, and it runs in your browser, so your text is never uploaded, stored or used for training. The report gives an AI-likeness score, sentence-by-sentence highlights and style notes with plainer wording, plus the hidden-character cleaner. To see what a classifier looks for, read how AI detectors work, or paste a text into the free AI detector.

Sources

1. OpenAI, Our approach to EU text provenance rules (5 October 2026), checked November 2026
2. OpenAI, textGrain: Entropy-Calibrated Watermarking for Language Model Text (technical report, 5 October 2026), checked November 2026
3. OpenAI, Verify content generated by OpenAI, checked November 2026
4. Anthropic, Claude text watermarking (14 August 2026, updated 1 September 2026), checked November 2026
5. Claude Help Center, How Claude marks AI-generated content, checked November 2026
6. Anthropic, Check if a file was made with Claude, checked November 2026
7. Google DeepMind, Watermarking AI-generated text and video with SynthID (14 May 2024), checked November 2026
8. Google AI for Developers, SynthID Text, checked November 2026
9. Google, Google expands SynthID Detector for AI content (7 October 2026), checked November 2026
10. Regulation (EU) 2024/1689 (AI Act), Article 50, checked November 2026
11. Regulation (EU) 2026/1744 (Digital Omnibus on AI), checked November 2026
12. European Commission, Code of Practice on Transparency of AI-Generated Content (10 June 2026), checked November 2026
13. European Commission, Strong backing for the Code of Practice (31 July 2026), checked November 2026