AI text watermarks: Claude, Gemini and OpenAI mark what they write

Since this summer, text from Claude and Gemini carries an invisible mark; since 5 October, so does text from ChatGPT and Codex in Europe, and OpenAI's API can be asked to add the mark. Text from Grok, Mistral, Meta AI and models run at home carries none, and Chinese services settle for a visible label. Yet outside the companies involved, nobody can check a text today.

Can a teacher tell that an essay came from an AI? Can a recruiter tell that a cover letter was written by ChatGPT? Since this summer, the answer depends on the model. Three companies now slip an invisible mark, a watermark, into the texts they write — Anthropic everywhere, Google everywhere, OpenAI in Europe — and their own tools can find it again. The others do not, or not yet.

This article takes stock as of 28 October 2026: who marks, which models, what the mark actually tells you, and what language changes. It was written by Claude Opus 5.5, a model made by Anthropic, one of the companies involved, so it carries that watermark itself.

Europe has required marking since 2 August 2026

Who marks AI-written text, and since when

Text output of each service, as of 28 October 2026.

2024202520262027China: labels requiredEU: AI Act art. 50EU deadline, older servicesShared detection (EU)28 Oct 2026Google · Gemini appsince May 2024Google · Gemini APIsince August 2026Anthropic · Claudesince August 2026Chinese apps, in Chinasince September 2025Mistralannounced for 2 Dec 2026OpenAI · ChatGPT, CodexEU + API since Oct 2026Meta · Muse, Llamano text watermarkSpaceX · Grokno text watermarkOpen models run locallyno text watermark
  • Invisible watermark in the text
  • Visible “AI” label (app, exported files)
  • Announced
View as a data table
ServiceText marking
Google · Gemini appsince May 2024
Google · Gemini APIsince August 2026
Anthropic · Claudesince August 2026
Chinese apps, in Chinasince September 2025
Mistralannounced for 2 Dec 2026
OpenAI · ChatGPT, CodexEU + API since Oct 2026
Meta · Muse, Llamano text watermark
SpaceX · Grokno text watermark
Open models run locallyno text watermark

Sources: provider announcements and documentation, AI Act (EU), CAC labelling measures (China); list at the end of the article.

PNGaiwritingbenchmark.com

The push comes from Article 50 of the AI Act, which requires providers of generative AI to make their output detectable by a machine, text included. An image can hide that information in its file; pasted text keeps nothing. So the words themselves have to be marked: the model picks its words along a statistical pattern that no reader notices but that software holding the right key can recognise.

The code of practice that comes with the law sets the bar at 200 tokens, a good hundred words; below that, a mark would be too weak to be reliable and is not required. Providers must also offer a free checking tool, which none has yet opened to the public.

Model by model

The table covers the models in our ranking.

Models in the ranking Mark in the text Who can read it
Claude Opus 5.5, Sonnet 5.5, Fable 5.1, Haiku 5.5 and earlier versions Invisible watermark, worldwide, in the app, the API and Claude Code Anthropic, plus approved organisations: regulators, media, researchers, schools and universities
Gemini 4 Argon, Gemini 3.8 Flash and earlier Flash versions Invisible watermark (SynthID), in the app since 2024 and in the API Google; its checking tool is still waitlisted
GPT-6.1 Sol and Terra, GPT-6 Sol, Luna and Astra, GPT-5.6 Since 5 October: invisible watermark (textGrain) in the EU (ChatGPT and Codex, all plans) and, in the API worldwide, on request OpenAI; the detector is by application (approved researchers and expert organisations)
Muse Spark 1.3 and 1.2 (Meta), Grok 4.7 and 4.6 (SpaceX Super Intelligence, formerly xAI), Nemotron (NVIDIA) and the other models in the ranking (Beam, Inkling, Aion, Agnes, Mercury) None —
Mistral Medium 3.5, Mistral Large 4 None, promised for 2 December —
DeepSeek, Qwen, GLM, Kimi (K3, K3.1 Pro), MiniMax, MiMo, Seed, Hy, Ling, Step Visible “AI” label in the apps, in China Anyone, on screen
Open models run locally (Gemma, Qwen, Granite, Nemotron, Muse Glimmer, Laguna…) None —

SpaceX stands apart: it is the only major Western provider that has not signed the European code, even though it folded xAI into the company in February 2026 and rebranded the Grok division since. The law still binds it.

OpenAI: Europe first, the API on request

On 5 October 2026, OpenAI published its approach to the EU text-provenance rules, with the technical report behind it. Its watermark, textGrain, is OpenAI’s own method rather than Google’s SynthID: it bends the model’s word choices the same way, and its own detector looks for the signal. Three things changed at once. In the European Union, eligible ChatGPT and Codex users on every plan get watermarked text — OpenAI called the rollout “the coming weeks” and has given no completion date, and it is deliberately not a global default at launch. In the API, customers anywhere can switch the mark on per organisation or project, for select models; it stays off by default. And a detector exists, but access went first to approved researchers and expert organisations — OpenAI named Cornell, ETH Zurich and the Slovak institute KInIT — with an open-source release planned, without a date.

OpenAI’s own numbers set the scale of what such a mark can and cannot do. At a 1% false-positive rate, the detector found textGrain in about 80% of 200-token passages and about 95% of 400-token passages on flexible subjects, much less on mathematics. In a test on 400-token passages, replacing 10% of the words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. OpenAI judged the false alarms and the misses too costly for a public tool at launch.

For this benchmark the practical consequence is simple: its GPT models are called through an API route where the mark is off by default, so the GPT responses in the ranking do not carry it.

What the mark actually tells you

For whoever receives a text, the mark is of little use today. No provider offers a public tool: Google’s checker is still waitlisted, Anthropic’s access goes to approved organisations, and OpenAI’s detector went first to approved researchers and expert organisations. A teacher therefore cannot check anything alone, unless their school is granted the access Anthropic offers.

The mark is also fragile. It survives light edits, but a full rewrite, a paraphrase by another model or a translation by another tool erases it. Nor does it say who wrote the text: Claude may have drafted the whole letter or merely fixed three sentences of a human draft, and the mark looks the same.

Finally, the detectors sold to schools and companies do not read these watermarks. They are classifiers that estimate a probability from style, with known failures: a 2023 study found they flagged about 61% of essays by non-native English speakers as AI-written. A detector score is not proof.

Chinese models: a label on screen, in China

China got there first by another route. Since 1 September 2025, its rules require chat and writing services to show an “AI” label in the interface and to write one into exported files. DeepSeek complied on day one; in September 2026 the regulator told Doubao, Yuanbao, Qwen and Wenxin to apply it more strictly.

For text, that label stays on the screen: a passage copied from the chat window leaves without it. No Chinese provider has announced a hidden watermark in the words, and nothing is planned for their APIs or their international versions, the ones used from Europe or the US. The newest arrival in our ranking confirms the pattern: Kimi K3.1 Pro is called through an international API route, where no label or mark is announced either.

Models run at home: no mark

The watermark is applied on the provider’s servers, at the moment the model picks each word. A model running on your own computer, like those on our Local AI page, therefore carries none, whoever made it. The only exception is deliberate: whoever installs the model can switch on the marking tool Google published, with their own key.

Does the language matter?

Not for the obligation: the law targets services offered in Europe whatever the language of the text, and Claude’s watermark applies everywhere. An English text from Claude or Gemini is marked like a French or German one.

Language matters for how the mark gets lost. A text written in another language and translated into English by a different tool loses its mark, while a translation made by Claude carries Claude’s. No provider has published detection rates by language. Detectors differ too: according to its documentation as quoted by third parties, Turnitin scores English, Spanish and Japanese, but not French or German, which matters for students who write in more than one language.

Sources

Conclusion

As of 28 October 2026, text written with Claude or Gemini carries an invisible mark worldwide; in Europe, text from ChatGPT and Codex carries one too, and OpenAI's API adds the mark on request. Text written with any other model does not. The mark proves little to whoever looks for it: only the makers can read it, it vanishes once the text is translated or rewritten with another tool, and it shows that a model took part, not that it wrote everything. OpenAI measured its own mark: at a 1% false-positive rate the detector finds it in about 80% of 200-token passages, and replacing a quarter of the words with synonyms cut detection from 92% to 17%. The detectors sold to schools do not read these marks: they guess. Two dates will change the picture: 2 December 2026 for services already available in Europe, 2 February 2027 for shared detection.