AI text watermarks compared: Claude, Gemini, ChatGPT and Llama (2026)

Two kinds of "watermark"

When a vendor says it watermarks AI content, it can mean two very different things. Metadata marks (C2PA content credentials, IPTC fields) are attached to a file: they are easy to verify and trivially lost when the file is screenshotted, re-encoded or copied as text. Statistical marks are embedded in the content itself by biasing the generator's choices; for text, that means word choice. They survive copy-paste and need a key to read. Everything below is about the second kind, applied to text, because that is the one that follows your words around.

Claude (Anthropic)

Status, August 2026: deployed. Anthropic said on August 11, 2026 that Claude "embeds an imperceptible watermark directly into the text," biasing word choices according to a key it holds. It applies to models launched on or after August 2, 2026, worldwide, across the app, Claude Code, Cowork, Tag and the API, with older models to follow. Files get C2PA metadata in addition. Detection tools are "forthcoming"; no thresholds published. Full details in Claude's text watermark, explained.

Gemini (Google)

Status: deployed since 2024. Google DeepMind's SynthID-Text was described in Nature in October 2024 and rolled out in the Gemini app and API. It uses "tournament sampling": at each step, candidate tokens are scored by key-dependent functions and the winner is slightly favoured, which leaves a detectable statistical trace with little effect on quality. Google open-sourced the algorithm through Hugging Face Transformers so that other providers can use it, and offers a SynthID Detector portal to check text, images, audio and video generated by Google models. Robustness claims are similar to Anthropic's: survives light edits, weakened by heavy paraphrase and translation, weak on short or very low-entropy text (for example a list of capital cities has almost no room to bias).

ChatGPT (OpenAI)

Status: not deployed for text. OpenAI has stated since 2023 that it has a text watermarking method and, in 2024, that it was "highly accurate" but had not been released, citing risks to non-native English writers and the ease of circumvention by paraphrase or translation. Forbes' August 2026 coverage, citing the Wall Street Journal, repeats that the capability exists and remains switched off for competitive reasons. OpenAI does attach C2PA metadata to images generated by DALL-E and GPT image models, and is a signatory of the EU transparency code, so this may change. As of this writing, ChatGPT text carries no statistical mark, which is why the "AI detectors" people run on it are all style-based guesswork.

Llama and open-weight models (Meta and others)

Status: no enforceable text watermark. Meta releases Llama weights openly; anyone running the model locally controls the sampling and could simply not apply a mark, so Meta does not claim one for text. Meta's provenance work targets media (Video Seal, Audio Seal, image credentials). The same logic applies to Mistral, Qwen, DeepSeek and every other open-weight family: a hosting provider could add a SynthID-style layer on its own endpoint, but the weights themselves cannot carry a mark. Text produced with these models is unmarked unless the specific service you used says otherwise.

Side by side

How to tell which mark a text carries

You usually cannot, from the text alone, and that is by design: a statistical mark is only readable with the vendor's key. What you can do is reason from provenance. If you know a text came out of a current Claude model, it is marked. If it came out of Gemini, it is marked with SynthID-Text. If it came out of ChatGPT or a locally run open model, there is no statistical mark to speak of, whatever a style detector says. When you do not know the provenance (a contributor's draft, a translation agency's delivery), the honest position in 2026 is: it may carry a Claude or a Gemini mark, and only Anthropic or Google could confirm it. This is also why Unmarker reports an estimated score computed from a model of the mechanism, and labels it as such, rather than pretending to hold a key it does not have.

What it means for you

• If you write with Claude or Gemini, assume the text is marked, including short "polish this" passes on your own draft. Only a deep rewrite or a translation changes that; see how to remove the Claude watermark.
• If you write with ChatGPT or an open model, there is no statistical mark to remove today; what flags your text is style, and that is a different fight.
• Regardless of vendor, a mark says "processed by this model", not "written by this model". Disclosure rules, where they apply to you, are about your process, not about the presence of a signal.
• Unmarker is built for the Claude mark first because it is the newest, it applies worldwide, and it catches the most people by surprise. Support for other deployed, measurable marks will follow. Join the early access list if that is your case.

Status as of August 15, 2026, from the vendors' public documentation and the press coverage listed below. Vendors change policies quickly; when a statement above no longer matches a vendor's documentation, the documentation wins.