AI Watermarks in Claude, ChatGPT and Gemini, Explained

The short version

Claude has marked text since September 2026, worldwide and only in newer models. Image, video and audio files Claude creates through code also get provenance data (C2PA).
ChatGPT marks images and audio, and, according to OpenAI, text only in the EU, rolled out gradually since October 2026.
Gemini marks images, video and audio with SynthID. According to Google, that includes text from the Gemini app.
You can check yourself only images, video, audio or files. None of the three providers offers a public check for text.
A match proves little: according to the providers, no watermark does not mean “made by a person”, and a watermark says nothing about authorship.

How a text watermark works

A language model writes word by word, or more precisely in tokens, which are pieces of words. For every position it has several fitting candidates with different probabilities. That is where the watermark comes in. The three graphics show it in simplified form, based on SynthID Text by Google DeepMind, published in the journal Nature (Dathathri et al., 2024); according to Anthropic, Claude's text watermark is a version of it.

Let’s move the meeting to

Tuesday40%

Wednesday35%

tomorrow25%

1. Word by word. For the next position the model has several fitting words, each with a probability. Without a watermark it draws one according to these probabilities. The numbers are a made-up example.

Let’s move the meeting to

Tuesday

Wednesdayfavored

tomorrow

I’ll send you the invoice by

Tuesday

Wednesday

tomorrowfavored

2. The key nudges slightly. A secret key and the preceding words decide, at every position, which of the fitting words is slightly favored. In the next sentence it is a different one. There is no fixed word list, and the text still reads normally.

Without watermark10 of 30

With watermark19 of 30

Detection according to OpenAI's own tests:

Psychology texts, about 200 tokens80%

Psychology texts, about 400 tokens95%

About 400 tokens, 25% of words replaced with synonyms17%

3. The detector counts. With the key, it checks across the whole text how often the favored words appear. Without a watermark only some land there by chance, with a watermark clearly more. In a few words that is still chance; over hundreds of words it becomes a clear signal. Rewriting with synonyms weakens it. The strips are a made-up example. The percentages were published by OpenAI for its own text watermark: the first two at a 1% false positive rate, the last from a second test in which unedited texts were detected about 92% of the time.

Nothing is added to the text, not even hidden characters; Anthropic and OpenAI both say so. According to Google, response quality stays the same, but the watermark is weaker on factual answers and after thorough rewriting or translation (Google DeepMind, Google).

What the three providers mark

Besides the invisible watermark in the content there is a second technique: Content Credentials (C2PA) are provenance data in the file that are lost with a screenshot or a conversion. As of October 22, 2026, according to the providers:

Claude

Text: watermarked since September 2026, worldwide. Anthropic announced it on August 14, 2026, and the first models that carry it from launch came out on September 1 (release notes). Anthropic is adding it to older models; which ones already have it is in a table on the help page that Anthropic keeps updating. On October 22, Mythos 5, Opus 4.5 to 4.7, Sonnet 4.5 and 4.6 and Haiku 4.5 did not have one yet.
What the watermark survives: according to Anthropic, light editing likely does not remove it, a full rewrite does. If Claude only proofreads your own text, it carries little watermark. Translations by Claude are marked. Code is marked less, because it has to be exact; comments in code can carry it.
Images, video, audio: according to the Transparency Hub, Claude has no image model of its own. Image, video and audio files Claude creates through code carry Content Credentials, PDF and Office files do not (docs).
Check yourself: files in the Claude Content Checker. Anthropic gives its text detector only in a private preview to eligible organizations, such as authorities, media, research and companies with legal obligations.

ChatGPT

Text: OpenAI developed a method in 2024 but did not ship it (OpenAI). Since October 2026 it has been rolling out a text watermark for ChatGPT and Codex in the EU, in response to the EU rules (OpenAI). Outside the EU, ChatGPT text is not marked; in the API, customers worldwide can switch it on for selected models. According to OpenAI, the method adds no hidden characters or spaces. How well its detector works is shown in graphic 3 above.
Images, video, audio: images carry Content Credentials and also SynthID, which OpenAI has used since May 2026 through a partnership with Google (OpenAI). Audio carries SynthID (OpenAI).
Check yourself: images and audio at openai.com/verify, which only finds OpenAI signals. The text detector is only available to approved researchers.

Gemini

Text: according to Google, text from the Gemini app is marked with SynthID (Google).
Images, video, audio: SynthID and Content Credentials. According to Google, SynthID is meant to survive cropping and compression. The visible mark can be switched off in the app, in India, South Korea and Vietnam only with AI Ultra; the invisible ones cannot (Google).
Check yourself: images, video and audio in the Gemini app or the SynthID Detector. Google names no check for text (Google).

At a glance

	Text	Images, video, audio, files	Check yourself
Claude	Watermark in newer models, including Opus 5.5, Opus 5, Sonnet 5.5 and Haiku 5.5, worldwide	No image model of its own. Images, video and audio Claude creates through code carry C2PA. PDF and Office files do not.	Files in the Claude Content Checker. Text only through a detector in private preview, for example for authorities, media and research.
ChatGPT	Watermark only in the EU, rolled out gradually since October 2026, for ChatGPT and Codex	Images with C2PA and SynthID, audio with SynthID	Images and audio at openai.com/verify. The text detector is only available to approved researchers.
Gemini	According to Google, text from the Gemini app is marked	Images, video and audio with SynthID and C2PA. The visible mark can be switched off in the app (in India, South Korea and Vietnam only with AI Ultra), the invisible ones cannot.	Image, video and audio in the Gemini app or the SynthID Detector. Google names no check for text.

Can you check for an AI watermark?

For images, video, audio and files yes, for text not publicly:

SynthID Detector: open to everyone since October 7, 2026. Checks image, video and audio from Google and partners such as OpenAI, NVIDIA and Kakao (Google).
Gemini app: upload a file and ask whether it was made with AI. Via SynthID it only recognizes Google content, but it also reads Content Credentials (Google).
OpenAI: openai.com/verify only finds OpenAI signals.
Claude Content Checker: checks files right in the browser, not text (Anthropic).

None of the three providers offers a public check for text (as of October 22, 2026): Anthropic and OpenAI give their text detectors only to selected parties, and Google names ways to check only for images, video and audio.

What a match means and what it doesn't

All three providers say the same: no watermark does not mean a person made the content. According to Anthropic, a detector often finds nothing: with older models that have no watermark, after heavy editing or translation, with very short texts, or when provenance data was lost through a screenshot or a conversion. In turn, a match does not prove the AI wrote the text: Claude may only have translated or summarized it. According to Anthropic and OpenAI, a watermark says nothing about authorship or liability.

How to remove watermarks is deliberately not covered here. According to the providers, a missing watermark proves nothing anyway.

What this means for your work

Texts Claude drafts for you carry a watermark with the newer models, according to Anthropic likely even after light editing. This applies worldwide, not only in the EU.
Texts from ChatGPT get a watermark in the EU, rolled out gradually since October 2026 across all plans; not outside the EU (OpenAI).
Images on social media: Instagram and Facebook show the “AI info” label when they detect provenance data (Meta). LinkedIn shows a Content Credentials icon (LinkedIn). YouTube can label videos with C2PA data itself, and you cannot change that label (YouTube).
An invisible watermark is not a notice for people. Whoever reads your text or looks at your image does not notice it.
If someone holds a text against you as “AI-detected”: the text detectors from Anthropic and OpenAI are not public. Anthropic says a match is not conclusive, and according to both, a watermark says nothing about authorship.

How to spot AI-written text without a tool, and how to keep your texts sounding like you, is in the guide How to Make AI Writing Sound Like You.

Labeling obligations: what the EU says

The background is Article 50 of the EU AI Act, which has applied since August 2, 2026. Its paragraph 2 is addressed to the providers of AI systems. What the article means for your specific case is not something I can judge; I am an AI consultant, not a lawyer. Here is what the responsible bodies say:

“Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.”

Article 50(2) AI Act, EUR-Lex

“Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, that have been placed on the market before 2 August 2026 shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026.”

Transitional period added by Regulation (EU) 2026/1744, EUR-Lex

“Deepfakes and AI-generated or AI-manipulated text published on matters of public interest must be clearly labelled.”

European Commission, news of June 10, 2026

“This obligation shall not apply […] where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.”

Article 50(4) AI Act on text of public interest, EUR-Lex

“This Regulation does not apply to obligations of deployers who are natural persons using AI systems in the course of a purely personal non-professional activity.”

Article 2(10) AI Act, EUR-Lex

“Any activity through which natural persons gain an economic benefit on a regular basis or are otherwise involved in a professional, business, trade, occupational or freelance activity should be considered as a ‘professional’ activity.”

European Commission, guidelines on Article 50, paragraph 19 (non-binding, according to the Commission)

For advice on your case, ask a lawyer or the authority responsible for the AI Act in the EU country where you work. The guidelines themselves state: “Any authoritative interpretation of the AI Act may ultimately only be given by the Court of Justice of the European Union (‘CJEU’).”

Frequently asked questions

Is ChatGPT text watermarked?

In the EU, yes, according to OpenAI: since October 2026, OpenAI has been gradually adding an invisible watermark to ChatGPT and Codex text there. Outside the EU, ChatGPT text is not marked; in the API, customers worldwide can switch it on for selected models. Images from ChatGPT carry Content Credentials (C2PA) and a SynthID watermark, audio carries SynthID. According to OpenAI, the text method adds no hidden characters.

Since when does Claude have a watermark?

Anthropic announced the text watermark on August 14, 2026. The first models with the watermark from launch came out on September 1, 2026, and older models are being updated. According to Anthropic, it applies worldwide, not only in the EU.

Which Claude models have a watermark?

According to Anthropic's help page of October 22, 2026, among others Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5, Haiku 5.5, Fable 5.1, Fable 5 and Mythos 5.1. Mythos 5, Opus 4.5 to 4.7, Sonnet 4.5 and 4.6 and Haiku 4.5 only had Content Credentials for files on that day. The list keeps changing.

Is there a Claude watermark detector?

Not a public one for text. Anthropic gives its text detector only in a private preview to eligible organizations, such as authorities, media, research and companies with legal obligations; OpenAI gives its own only to approved researchers. Files made with Claude can be checked by anyone in the Claude Content Checker, and images, video and audio from Google and partners in the SynthID Detector.

Do AI images have a watermark?

Images from Gemini and ChatGPT carry an invisible SynthID watermark and Content Credentials. The provenance data is lost with a screenshot or conversion; according to Google, SynthID is meant to survive cropping and compression. Claude has no image model of its own; images it creates through code carry Content Credentials.

Do I have to label AI content?

I can't judge that for you; I'm not a lawyer. The European Commission writes: “Deepfakes and AI-generated or AI-manipulated text published on matters of public interest must be clearly labelled.” Article 2(10) of the AI Act reads: “This Regulation does not apply to obligations of deployers who are natural persons using AI systems in the course of a purely personal non-professional activity.” For advice on your case, ask a lawyer or the authority responsible for the AI Act in your EU country.

Sources

Retrieved on October 22, 2026.

Want to know which of your tasks AI can take on?

I work 1:1 with freelancers, consultants, coaches and small teams on their real tasks. The intro call is free, takes 15 minutes and is not a sales pitch.

Book Free Consultation

AI watermarks in Claude, ChatGPT and Gemini