textGrain vs SynthID vs the Claude watermark
Three of the biggest AI companies now add watermarks to text. They share one basic idea. They differ in where it is turned on, who can check it, and how much they have shared.
All three hide a pattern in word choices using a secret key. None of them adds visible marks or characters. Google’s SynthID Text is the oldest and is open source. Reports say Anthropic built the Claude watermark on the SynthID method. OpenAI’s textGrain is the newest. It starts in the EU, and OpenAI says it did as well as or better than the others it tested.
Side by side
| textGrain | SynthID Text | Claude watermark | |
|---|---|---|---|
| Company | OpenAI | Google DeepMind | Anthropic |
| Announced | 5 Oct 2026 | Published in Nature, Oct 2024 | August 2026 |
| How it works | A secret key gives small pushes to word choices | Picks between word options in rounds, using a key | Reports say it is based on SynthID Text |
| Where it is on | ChatGPT and Codex in the EU. Optional in the API everywhere | Google’s Gemini products | Reported for new Claude models |
| Who can detect it | Approved researchers and expert groups | SynthID Detector portal for chosen testers | A private test for chosen groups, as reported |
| Open source | Planned, no date yet | Yes, in Hugging Face Transformers since Oct 2024 | Not announced |
| Shared test results | 92% → 66% → 17% with 0, 10 and 25% of words changed (400 tokens) | Weaker on short and heavily edited text, says Google’s paper | Not shared in detail |
What they have in common
All three add the watermark while the text is written. They change how the model picks words. They do not stamp anything onto finished text. So they share three traits:
- Invisible, with no drop in quality, at least in the companies’ own tests.
- Locked to a key. Only the owner of the key can test for the pattern.
- They need long, unedited text. Short or heavily rewritten text carries too little signal.
Our watermark demo shows this shared idea in a few seconds.
Where they differ
Where it is turned on
OpenAI chose to start with the EU only and learn from real use first. Reports say Anthropic uses its watermark more widely on new Claude models. Google uses SynthID across its own products.
Openness
Google made SynthID Text open source. Anyone can use it to watermark their own model’s text, but not to detect Google’s. OpenAI says it will make textGrain open source later. Anthropic says it plans to share more technical details.
Who can detect it
All three limit who can use their detector. As of October 2026, none of them offers a free public checker for text.
What about images?
Text watermarks are separate from image marks. OpenAI and Google images carry C2PA content credentials, and Google also adds SynthID to images, audio and video. Removing data from a file does nothing to a text watermark, and the other way round.
Questions people ask
Which text watermark is the strongest?
No independent test has compared them yet. OpenAI says textGrain did as well as or better than the other methods it tested, including SynthID Text. But that is OpenAI’s own test.
Can one company detect another company’s watermark?
No. Each detector only knows the pattern made with its own key.
Is there one checker for all AI text watermarks?
No. Each company controls its own detector and who can use it. A tool that says it can check them all from pasted text is guessing.